$ docker run -d --name lynkr -p 8081:8081 -e MODEL_PROVIDER=databricks lynkr:latest

Updatedlast week

LanguageJavaScript

Setup difficultyEasy

Stars: 639
RAM: 8GB+ RAM for 7-8B models, 16GB+ RAM for 14B models (local model hosting)
Ports: 8081 (HTTP proxy/gateway and built-in dashboard (configurable via PORT)), 8787 (optional Headroom context-compression sidecar container)
Requires: Node.js

What is Lynkr?

Lynkr is a Node.js gateway that fronts AI coding assistants — Claude Code, Cursor, the Codex CLI, Cline and Continue.dev — and answers both the Anthropic-style and the OpenAI-compatible endpoint formats, so clients need no code changes. On the way to the model it binary-compresses large JSON tool results (87.6% fewer tokens on a 60-item grep array in its own benchmark), drops tool schemas a request will not use, and answers near-duplicate prompts from a semantic cache in roughly 171ms with nothing billed. Each request can be scored for difficulty and sent to one of four tiers configured as provider:model pairs, keeping easy turns on a free local Ollama model while hard ones escalate to cloud capacity; sessions stay pinned to one model until a guard forces a re-route. A dashboard served under /dashboard reports spend, counterfactual savings, tier mix, routing accuracy, logs and provider health, alongside Prometheus metrics, health endpoints and a no-restart admin reload. Installation is a global npm package (also a Homebrew tap, a curl installer, Docker Compose or a source build), configuration is an interactive lynkr init wizard or a .env file, and the gateway listens on port 8081 with local SQLite files for its data. It is Apache 2.0 licensed, free, and the project itself notes that several provider entries are documented but not yet probed against live keys.

Should you self-host Lynkr?

Lynkr scores 48/100 on selfhost.directory (how scoring works) and ranks #59 of 125 in AI. It has 639 GitHub stars.

There was no stable release in the past 12 months, so check the repository's recent activity before you rely on it.

Setup is easy: most people get it running in one sitting. It is Apache-2.0-licensed and written mostly in JavaScript.

Its closest alternative on selfhost.directory is OmniRoute (88/100, 82 107 stars). It scores higher overall, so compare the two before you pick. See Lynkr vs OmniRoute side by side.

Features

Built-in Dashboard
Complexity tier routingAll four TIER_* variables must be set for tier routing to activate.
Cost tracking with model pricing
Health check endpoints
Interactive setup wizard
Load shedding and fallback
Long-term MemoryDisabled by default; enabled with MEMORY_ENABLED=true.
MCP Support
Prometheus Metrics
Semantic Caching
SSE streaming
Sticky Sessions
Token compression of tool results
Tool schema pruning
Wrap mode for AI coding tools

How to configure Lynkr

AGENTS_ENABLEDtrueDESCRIPTION: Inject agent-delegation instructions into the system prompt when the client declares a Task tool (tools always execute on the client). Values: true | false
DESCRIPTION: Inject agent-delegation instructions into the system prompt when the client declares a Task tool (tools always execute on the client). Values: true | false
ANTHROPIC_API_KEYAnthropic (direct + OAuth subscription mode) OAuth mode: when using `lynkr wrap claude`, the OAuth token from `claude login` is forwarded automatically — no API key needed. DESCRIPTION: Anthropic API key. Only needed when NOT using OAuth subscription.
Anthropic (direct + OAuth subscription mode) OAuth mode: when using `lynkr wrap claude`, the OAuth token from `claude login` is forwarded automatically — no API key needed. DESCRIPTION: Anthropic API key. Only needed when NOT using OAuth subscription.
ATLASCLOUD_API_KEYAtlas Cloud (OpenAI-compatible) DESCRIPTION: Atlas Cloud API key from https://www.atlascloud.ai/console/api-keys.
Atlas Cloud (OpenAI-compatible) DESCRIPTION: Atlas Cloud API key from https://www.atlascloud.ai/console/api-keys.
ATLASCLOUD_ENDPOINThttps://api.atlascloud.ai/v1/chat/completionsDESCRIPTION: Atlas Cloud Chat Completions endpoint.
DESCRIPTION: Atlas Cloud Chat Completions endpoint.
ATLASCLOUD_MODELqwen/qwen3.8-maxDESCRIPTION: Default Atlas Cloud text model.
DESCRIPTION: Default Atlas Cloud text model.
AWS_BEDROCK_API_KEYAWS Bedrock (uses Bedrock API Key, NOT IAM) DESCRIPTION: Bedrock bearer API key (starts with ABSK, generated in AWS console).
AWS Bedrock (uses Bedrock API Key, NOT IAM) DESCRIPTION: Bedrock bearer API key (starts with ABSK, generated in AWS console).
AWS_BEDROCK_MODEL_IDus.anthropic.claude-3-5-sonnet-20241022-v2:0DESCRIPTION: Default Bedrock model id (often a US inference profile).
DESCRIPTION: Default Bedrock model id (often a US inference profile).
AWS_BEDROCK_REGIONus-east-1DESCRIPTION: AWS region.
DESCRIPTION: AWS region.
AWS_REGIONDESCRIPTION: Fallback region if AWS_BEDROCK_REGION is unset.
DESCRIPTION: Fallback region if AWS_BEDROCK_REGION is unset.
AZURE_ANTHROPIC_API_KEYDESCRIPTION: API key for Azure Anthropic. Not needed if OAuth token is being forwarded.
DESCRIPTION: API key for Azure Anthropic. Not needed if OAuth token is being forwarded.
AZURE_ANTHROPIC_ENDPOINThttps://api.anthropic.com/v1/messagesAzure Anthropic (Anthropic-format endpoint, OAuth-friendly) DESCRIPTION: Endpoint URL for Azure-hosted Anthropic (also used as OAuth passthrough target).
Azure Anthropic (Anthropic-format endpoint, OAuth-friendly) DESCRIPTION: Endpoint URL for Azure-hosted Anthropic (also used as OAuth passthrough target).
AZURE_ANTHROPIC_VERSION2023-06-01DESCRIPTION: Anthropic API version header.
DESCRIPTION: Anthropic API version header.
AZURE_OPENAI_API_KEYDESCRIPTION: Azure OpenAI API key.
DESCRIPTION: Azure OpenAI API key.
AZURE_OPENAI_API_VERSION2024-08-01-previewDESCRIPTION: API version to use.
DESCRIPTION: API version to use.
AZURE_OPENAI_DEPLOYMENTgpt-4oDESCRIPTION: Deployment name (e.g. gpt-4o, gpt-5.2-chat).
DESCRIPTION: Deployment name (e.g. gpt-4o, gpt-5.2-chat).
AZURE_OPENAI_ENDPOINTAzure OpenAI DESCRIPTION: Azure OpenAI endpoint URL (standard or AI Foundry format).
Azure OpenAI DESCRIPTION: Azure OpenAI endpoint URL (standard or AI Foundry format).
BAIDU_API_KEYBaidu Qianfan (ERNIE) DESCRIPTION: Baidu Qianfan API key (format: bce-v3/ALTAK-...).
Baidu Qianfan (ERNIE) DESCRIPTION: Baidu Qianfan API key (format: bce-v3/ALTAK-...).
BAIDU_ENDPOINThttps://qianfan.baidubce.com/v2/chat/completionsDESCRIPTION: Qianfan v2 OpenAI-compatible chat completions endpoint.
DESCRIPTION: Qianfan v2 OpenAI-compatible chat completions endpoint.
BAIDU_MODELglm-5.2DESCRIPTION: Default Baidu ERNIE model.
DESCRIPTION: Default Baidu ERNIE model.
CAVEMAN_ENABLEDDESCRIPTION: Caveman terse-output injection (cuts output tokens at the cost of style). Values: true | false
DESCRIPTION: Caveman terse-output injection (cuts output tokens at the cost of style). Values: true | false
CAVEMAN_LEVELliteDESCRIPTION: Aggressiveness of the brevity instruction. Values: lite | full | ultra
DESCRIPTION: Aggressiveness of the brevity instruction. Values: lite | full | ultra
CLUSTER_ENABLEDfalseDESCRIPTION: Run the proxy in multi-worker cluster mode. Values: true | false
DESCRIPTION: Run the proxy in multi-worker cluster mode. Values: true | false
CLUSTER_WORKERSautoDESCRIPTION: Worker count. Values: auto | <integer>
DESCRIPTION: Worker count. Values: auto | <integer>
CODE_GRAPH_COMMANDgraphifyDESCRIPTION: External indexer binary the code-graph feature shells out to.
DESCRIPTION: External indexer binary the code-graph feature shells out to.
CODE_GRAPH_ENABLEDfalseDESCRIPTION: Master switch for the code-graph indexer. Values: true | false
DESCRIPTION: Master switch for the code-graph indexer. Values: true | false
CODE_GRAPH_TIMEOUT10000DESCRIPTION: Per-call timeout (ms).
DESCRIPTION: Per-call timeout (ms).
CODE_GRAPH_WORKSPACEDESCRIPTION: Workspace path the indexer operates on (defaults to cwd).
DESCRIPTION: Workspace path the indexer operates on (defaults to cwd).
CODEX_BINARY_PATHDESCRIPTION: Path to the `codex` binary; auto-detected if unset.
DESCRIPTION: Path to the `codex` binary; auto-detected if unset.
CODEX_ENABLEDCodex (uses your ChatGPT subscription via local codex CLI) DESCRIPTION: Enable the Codex local provider (requires `codex` CLI installed). Values: true | false
Codex (uses your ChatGPT subscription via local codex CLI) DESCRIPTION: Enable the Codex local provider (requires `codex` CLI installed). Values: true | false
CODEX_MODELDESCRIPTION: Codex model id.
DESCRIPTION: Codex model id.
CODEX_TIMEOUTDESCRIPTION: Per-request timeout in ms.
DESCRIPTION: Per-request timeout in ms.
CURSOR_AUTO_APPROVEDESCRIPTION: Auto-approve ALL cursor-agent permission requests, shell execution included (passes --force). The sandbox cwd is not a security boundary while this is on. Set false to approve MCPs only and reject edit/execute/delete requests. Values: true | false
DESCRIPTION: Auto-approve ALL cursor-agent permission requests, shell execution included (passes --force). The sandbox cwd is not a security boundary while this is on. Set false to approve MCPs only and reject edit/execute/delete requests. Values: true | false
CURSOR_BINARY_PATHDESCRIPTION: Path to the `cursor-agent` binary; auto-detected if unset.
DESCRIPTION: Path to the `cursor-agent` binary; auto-detected if unset.
CURSOR_ENABLEDCursor (uses your Cursor Pro subscription via local cursor-agent CLI) DESCRIPTION: Enable the Cursor local provider (requires `cursor-agent` CLI installed + `agent login`). Values: true | false
Cursor (uses your Cursor Pro subscription via local cursor-agent CLI) DESCRIPTION: Enable the Cursor local provider (requires `cursor-agent` CLI installed + `agent login`). Values: true | false
CURSOR_MODELDESCRIPTION: Cursor model id (run `cursor-agent models` to list).
DESCRIPTION: Cursor model id (run `cursor-agent models` to list).
CURSOR_TIMEOUTDESCRIPTION: Per-request timeout in ms.
DESCRIPTION: Per-request timeout in ms.
DATABRICKS_API_BASEDatabricks DESCRIPTION: Databricks workspace base URL.
Databricks DESCRIPTION: Databricks workspace base URL.
DATABRICKS_API_KEYDESCRIPTION: Databricks personal access token.
DESCRIPTION: Databricks personal access token.
DATABRICKS_ENDPOINT_PATHDESCRIPTION: Path to the serving endpoint to invoke.
DESCRIPTION: Path to the serving endpoint to invoke.
DEBUG_SQLDESCRIPTION: Pretty-print SQL statements to stdout (very verbose). Values: 1 | unset
DESCRIPTION: Pretty-print SQL statements to stdout (very verbose). Values: 1 | unset
EDENAI_API_KEYEden AI (600+ models via single API, OpenAI-compatible, EU/GDPR) DESCRIPTION: Eden AI API key.
Eden AI (600+ models via single API, OpenAI-compatible, EU/GDPR) DESCRIPTION: Eden AI API key.
EDENAI_EMBEDDINGS_MODELopenai/text-embedding-ada-002DESCRIPTION: Embedding model used through Eden AI.
DESCRIPTION: Embedding model used through Eden AI.
EDENAI_ENDPOINThttps://api.edenai.run/v3/chat/completionsDESCRIPTION: Eden AI chat completions endpoint (OpenAI-compatible v3).
DESCRIPTION: Eden AI chat completions endpoint (OpenAI-compatible v3).
EDENAI_MODELopenai/gpt-4o-miniDESCRIPTION: Default Eden AI model (provider/model naming).
DESCRIPTION: Default Eden AI model (provider/model naming).
EMBEDDINGS_PROVIDEREmbeddings provider override DESCRIPTION: Force a specific embeddings provider (otherwise inferred from MODEL_PROVIDER). One of: ollama, llamacpp, openrouter, openai
Embeddings provider override DESCRIPTION: Force a specific embeddings provider (otherwise inferred from MODEL_PROVIDER). One of: ollama, llamacpp, openrouter, openai
FALLBACK_ENABLEDfalseDESCRIPTION: Auto-fallback when the tier provider fails. Auto-true under tier routing. Values: true | false
DESCRIPTION: Auto-fallback when the tier provider fails. Auto-true under tier routing. Values: true | false
FALLBACK_PROVIDERdatabricksDESCRIPTION: Fallback provider when tier provider fails (cannot be local). One of: databricks, azure-anthropic, azure-openai, openrouter, edenai, openai, atlas, bedrock
DESCRIPTION: Fallback provider when tier provider fails (cannot be local). One of: databricks, azure-anthropic, azure-openai, openrouter, edenai, openai, atlas, bedrock
FILES_MAX_COUNT1000DESCRIPTION: Max stored files (LRU cap).
DESCRIPTION: Max stored files (LRU cap).
FILES_MAX_SIZE_MB100DESCRIPTION: Max per-file size in MB.
DESCRIPTION: Max per-file size in MB.
FILES_STORAGE_PATH./data/filesDESCRIPTION: Storage path for the Files tool.
DESCRIPTION: Storage path for the Files tool.
FIREWORKS_API_KEYFireworks AI (serverless open models) DESCRIPTION: Fireworks API key (format: fw-...). Get one at fireworks.ai.
Fireworks AI (serverless open models) DESCRIPTION: Fireworks API key (format: fw-...). Get one at fireworks.ai.
FIREWORKS_ENDPOINThttps://api.fireworks.ai/inference/v1/chat/completionsDESCRIPTION: Fireworks OpenAI-compatible chat completions endpoint.
DESCRIPTION: Fireworks OpenAI-compatible chat completions endpoint.
FIREWORKS_MODELaccounts/fireworks/models/kimi-k2-instruct-0905DESCRIPTION: Default Fireworks model (long-form serverless id). Dated suffixes rotate — check https://app.fireworks.ai/models for current ids.
DESCRIPTION: Default Fireworks model (long-form serverless id). Dated suffixes rotate — check https://app.fireworks.ai/models for current ids.
FIREWORKS_REASONING_EFFORTFireworks / OpenAI-compatible: reasoning effort and schema forwarding
Fireworks / OpenAI-compatible: reasoning effort and schema forwarding
FIREWORKS_REASONING_EFFORT_MAP
FMT_GUARD_ENABLEDDisable the markdown format guard (contradicts clients that demand raw JSON).
Disable the markdown format guard (contradicts clients that demand raw JSON).
FORCE_TIER_PATTERNSRouting hygiene for agent harnesses / structured-output clients Disable keyword-driven force escalation (shouldForceCloud/shouldForceReasoning).
Routing hygiene for agent harnesses / structured-output clients Disable keyword-driven force escalation (shouldForceCloud/shouldForceReasoning).
GCF_ENABLEDfalseDESCRIPTION: Enable GCF (Graph Compact Format) encoding for large structured payloads. Drop-in alternative to TOON; takes precedence over TOON when enabled. Opt-in. Values: true | false
DESCRIPTION: Enable GCF (Graph Compact Format) encoding for large structured payloads. Drop-in alternative to TOON; takes precedence over TOON when enabled. Opt-in. Values: true | false
GCF_FAIL_OPENtrueDESCRIPTION: Continue without GCF on encoder failure instead of erroring. Values: true | false
DESCRIPTION: Continue without GCF on encoder failure instead of erroring. Values: true | false
GCF_LOG_STATStrueDESCRIPTION: Log per-request GCF savings stats. Values: true | false
DESCRIPTION: Log per-request GCF savings stats. Values: true | false
GCF_MIN_BYTES4096DESCRIPTION: Minimum byte size before GCF encoding kicks in.
DESCRIPTION: Minimum byte size before GCF encoding kicks in.
GCF_VERIFYtrueDESCRIPTION: Round-trip verify each encoding and keep the original JSON on any mismatch (makes the compression provably lossless per payload; adds one decode per converted blob). Values: true | false
DESCRIPTION: Round-trip verify each encoding and keep the original JSON on any mismatch (makes the compression provably lossless per payload; adds one decode per converted blob). Values: true | false
GOOGLE_API_KEYDESCRIPTION: Fallback Google API key if VERTEX_API_KEY is unset.
DESCRIPTION: Fallback Google API key if VERTEX_API_KEY is unset.
HEADROOM_CACHE_ALIGNERtrueDESCRIPTION: Cache-aligner transform (align cache breakpoints). Values: true | false
DESCRIPTION: Cache-aligner transform (align cache breakpoints). Values: true | false
HEADROOM_CCRtrueDESCRIPTION: CCR (Compress-Cache-Retrieve) mode. Values: true | false
DESCRIPTION: CCR (Compress-Cache-Retrieve) mode. Values: true | false
HEADROOM_CCR_TTL300DESCRIPTION: TTL for CCR cached chunks (seconds).
DESCRIPTION: TTL for CCR cached chunks (seconds).
HEADROOM_DOCKER_AUTO_BUILDtrueDESCRIPTION: Auto-build the image if not found locally. Values: true | false
DESCRIPTION: Auto-build the image if not found locally. Values: true | false
HEADROOM_DOCKER_BUILD_CONTEXT./headroom-sidecarDESCRIPTION: Build context path when auto-building the sidecar image.
DESCRIPTION: Build context path when auto-building the sidecar image.
HEADROOM_DOCKER_CONTAINER_NAMElynkr-headroomDESCRIPTION: Sidecar container name.
DESCRIPTION: Sidecar container name.
HEADROOM_DOCKER_CPU_LIMIT1.0DESCRIPTION: Sidecar CPU limit.
DESCRIPTION: Sidecar CPU limit.
HEADROOM_DOCKER_ENABLEDtrueDESCRIPTION: Auto-manage a Docker container for the sidecar. Values: true | false
DESCRIPTION: Auto-manage a Docker container for the sidecar. Values: true | false
HEADROOM_DOCKER_IMAGElynkr/headroom-sidecar:latestDESCRIPTION: Sidecar image name.
DESCRIPTION: Sidecar image name.
HEADROOM_DOCKER_MEMORY_LIMIT512mDESCRIPTION: Sidecar memory limit.
DESCRIPTION: Sidecar memory limit.
HEADROOM_DOCKER_NETWORKDESCRIPTION: Optional Docker network the sidecar joins.
DESCRIPTION: Optional Docker network the sidecar joins.
HEADROOM_DOCKER_PORT8787DESCRIPTION: Host port the sidecar publishes on.
DESCRIPTION: Host port the sidecar publishes on.
HEADROOM_DOCKER_RESTART_POLICYunless-stoppedDESCRIPTION: Docker restart policy for the sidecar.
DESCRIPTION: Docker restart policy for the sidecar.
HEADROOM_ENABLEDtrueDESCRIPTION: Master switch for Headroom sidecar context compression. Values: true | false
DESCRIPTION: Master switch for Headroom sidecar context compression. Values: true | false
HEADROOM_ENDPOINThttp://localhost:8787DESCRIPTION: Headroom sidecar endpoint.
DESCRIPTION: Headroom sidecar endpoint.
HEADROOM_KEEP_TURNS10DESCRIPTION: How many recent turns the rolling window keeps verbatim.
DESCRIPTION: How many recent turns the rolling window keeps verbatim.
HEADROOM_LLMLINGUAfalseDESCRIPTION: LLMLingua ML compression (requires GPU). Values: true | false
DESCRIPTION: LLMLingua ML compression (requires GPU). Values: true | false
HEADROOM_LLMLINGUA_DEVICEautoDESCRIPTION: Device for LLMLingua model. Values: auto | cpu | cuda | mps
DESCRIPTION: Device for LLMLingua model. Values: auto | cpu | cuda | mps
HEADROOM_LOG_LEVELinfoDESCRIPTION: Sidecar log level. Values: debug | info | warning | error
DESCRIPTION: Sidecar log level. Values: debug | info | warning | error
HEADROOM_MIN_TOKENSDESCRIPTION: Skip compression below this estimated token count.
DESCRIPTION: Skip compression below this estimated token count.
HEADROOM_MODEoptimizeDESCRIPTION: Operating mode. Values: audit (observe only) | optimize (apply)
DESCRIPTION: Operating mode. Values: audit (observe only) | optimize (apply)
HEADROOM_PROVIDERanthropicDESCRIPTION: Provider hint that selects which cache markers to emit. Values: anthropic | openai | google
DESCRIPTION: Provider hint that selects which cache markers to emit. Values: anthropic | openai | google
HEADROOM_ROLLING_WINDOWtrueDESCRIPTION: Rolling-window transform (keep newest N turns intact). Values: true | false
DESCRIPTION: Rolling-window transform (keep newest N turns intact). Values: true | false
HEADROOM_SMART_CRUSHERtrueDESCRIPTION: Smart-crusher transform (compress large text blocks). Values: true | false
DESCRIPTION: Smart-crusher transform (compress large text blocks). Values: true | false
HEADROOM_SMART_CRUSHER_MAX_ITEMS15DESCRIPTION: Max items the smart-crusher processes per request.
DESCRIPTION: Max items the smart-crusher processes per request.
HEADROOM_SMART_CRUSHER_MIN_TOKENSDESCRIPTION: Min token count before smart-crusher engages.
DESCRIPTION: Min token count before smart-crusher engages.
HEADROOM_TIMEOUT_MS5000DESCRIPTION: Sidecar request timeout in ms.
DESCRIPTION: Sidecar request timeout in ms.
HEADROOM_TOOL_CRUSHERtrueDESCRIPTION: Tool-crusher transform (compress large tool results). Values: true | false
DESCRIPTION: Tool-crusher transform (compress large tool results). Values: true | false
HISTORY_COMPRESSION_ENABLEDtrueDESCRIPTION: Summarize older history turns instead of dropping them. Values: true | false
DESCRIPTION: Summarize older history turns instead of dropping them. Values: true | false
HISTORY_KEEP_RECENT_TURNS10DESCRIPTION: How many recent turns to keep verbatim.
DESCRIPTION: How many recent turns to keep verbatim.
HISTORY_SUMMARIZE_OLDERtrueDESCRIPTION: Summarize history older than the recent window. Values: true | false
DESCRIPTION: Summarize history older than the recent window. Values: true | false
HOT_RELOAD_DEBOUNCE_MS1000DESCRIPTION: Debounce window for the reload watcher (ms).
DESCRIPTION: Debounce window for the reload watcher (ms).
HOT_RELOAD_ENABLEDDESCRIPTION: Hot-reload config when .env changes. Values: true | false
DESCRIPTION: Hot-reload config when .env changes. Values: true | false
INJECT_TOOLS_LLAMACPPtrueDESCRIPTION: Inject native tool definitions into llama.cpp requests. Values: true | false (default true)
DESCRIPTION: Inject native tool definitions into llama.cpp requests. Values: true | false (default true)
INJECT_TOOLS_OLLAMAtrueDESCRIPTION: Inject native tool definitions into Ollama requests (for models without tool-calling). Values: true | false (default true)
DESCRIPTION: Inject native tool definitions into Ollama requests (for models without tool-calling). Values: true | false (default true)
LARGE_PAYLOAD_OPTIMIZATIONtrueDESCRIPTION: Large-payload optimization (chunked encoding, streaming). Values: true | false
DESCRIPTION: Large-payload optimization (chunked encoding, streaming). Values: true | false
LARGE_PAYLOAD_THRESHOLD1048576DESCRIPTION: Bytes that count as a "large" payload.
DESCRIPTION: Bytes that count as a "large" payload.
LLAMACPP_API_KEYDESCRIPTION: Optional bearer token if your llama.cpp server requires auth.
DESCRIPTION: Optional bearer token if your llama.cpp server requires auth.
LLAMACPP_EMBEDDINGS_ENDPOINThttp://localhost:8080/embeddingsDESCRIPTION: Embeddings endpoint (defaults to ENDPOINT + /embeddings).
DESCRIPTION: Embeddings endpoint (defaults to ENDPOINT + /embeddings).
LLAMACPP_ENDPOINThttp://localhost:8080llama.cpp (local GGUF server) DESCRIPTION: llama.cpp server base URL.
llama.cpp (local GGUF server) DESCRIPTION: llama.cpp server base URL.
LLAMACPP_MODELdefaultDESCRIPTION: llama.cpp model name to request.
DESCRIPTION: llama.cpp model name to request.
LLAMACPP_TIMEOUT_MS120000DESCRIPTION: Per-request timeout in ms.
DESCRIPTION: Per-request timeout in ms.
LLM_AUDIT_ANNOTATIONStrueDESCRIPTION: Include audit annotations on each entry. Values: true | false
DESCRIPTION: Include audit annotations on each entry. Values: true | false
LLM_AUDIT_DEDUP_CACHE_SIZE100DESCRIPTION: LRU cache size for the dedup dictionary.
DESCRIPTION: LRU cache size for the dedup dictionary.
LLM_AUDIT_DEDUP_DICT_PATHDESCRIPTION: Path to the dedup dictionary file.
DESCRIPTION: Path to the dedup dictionary file.
LLM_AUDIT_DEDUP_ENABLEDtrueDESCRIPTION: Deduplicate repeated audit payloads via dictionary compression. Values: true | false
DESCRIPTION: Deduplicate repeated audit payloads via dictionary compression. Values: true | false
LLM_AUDIT_DEDUP_MIN_SIZE500DESCRIPTION: Smallest payload size (bytes) eligible for dedup.
DESCRIPTION: Smallest payload size (bytes) eligible for dedup.
LLM_AUDIT_DEDUP_SANITIZEtrueDESCRIPTION: Sanitize secrets out of dedup-eligible payloads. Values: true | false
DESCRIPTION: Sanitize secrets out of dedup-eligible payloads. Values: true | false
LLM_AUDIT_DEDUP_SESSION_CACHEtrueDESCRIPTION: Cache dedup state per session. Values: true | false
DESCRIPTION: Cache dedup state per session. Values: true | false
LLM_AUDIT_ENABLEDfalseDESCRIPTION: Master switch for the LLM audit log. Values: true | false
DESCRIPTION: Master switch for the LLM audit log. Values: true | false
LLM_AUDIT_LOG_FILEDESCRIPTION: Audit log file path.
DESCRIPTION: Audit log file path.
LLM_AUDIT_MAX_CONTENT_LENGTH5000DESCRIPTION: Legacy fallback for max content length.
DESCRIPTION: Legacy fallback for max content length.
LLM_AUDIT_MAX_FILES30DESCRIPTION: Rotated audit file count.
DESCRIPTION: Rotated audit file count.
LLM_AUDIT_MAX_RESPONSE_LENGTH3000DESCRIPTION: Max response chars retained per audit entry.
DESCRIPTION: Max response chars retained per audit entry.
LLM_AUDIT_MAX_SIZE100MDESCRIPTION: Per-file rotation size.
DESCRIPTION: Per-file rotation size.
LLM_AUDIT_MAX_SYSTEM_LENGTH2000DESCRIPTION: Max system-prompt chars retained per audit entry.
DESCRIPTION: Max system-prompt chars retained per audit entry.
LLM_AUDIT_MAX_USER_LENGTH3000DESCRIPTION: Max user-message chars retained per audit entry.
DESCRIPTION: Max user-message chars retained per audit entry.
LMSTUDIO_API_KEYDESCRIPTION: Optional bearer token.
DESCRIPTION: Optional bearer token.
LMSTUDIO_ENDPOINThttp://localhost:1234LM Studio DESCRIPTION: LM Studio OpenAI-compatible endpoint.
LM Studio DESCRIPTION: LM Studio OpenAI-compatible endpoint.
LMSTUDIO_MODELdefaultDESCRIPTION: LM Studio model id to request.
DESCRIPTION: LM Studio model id to request.
LMSTUDIO_TIMEOUT_MS120000DESCRIPTION: Per-request timeout in ms.
DESCRIPTION: Per-request timeout in ms.
LOAD_SHEDDING_ACTIVE_REQUESTS_THRESHOLD1000DESCRIPTION: In-flight request count above which load shedding triggers.
DESCRIPTION: In-flight request count above which load shedding triggers.
LOAD_SHEDDING_HEAP_THRESHOLD0.95DESCRIPTION: Heap-utilization fraction above which load shedding triggers.
DESCRIPTION: Heap-utilization fraction above which load shedding triggers.
LOAD_SHEDDING_MEMORY_THRESHOLD0.85DESCRIPTION: RSS-memory fraction above which load shedding triggers.
DESCRIPTION: RSS-memory fraction above which load shedding triggers.
LOG_FILE_ENABLEDDESCRIPTION: Persistent file logging via pino-roll. Values: true | false
DESCRIPTION: Persistent file logging via pino-roll. Values: true | false
LOG_FILE_FREQUENCYDESCRIPTION: Roll frequency. Values: daily | hourly | <bytes>
DESCRIPTION: Roll frequency. Values: daily | hourly | <bytes>
LOG_FILE_LEVELDESCRIPTION: Log file verbosity.
DESCRIPTION: Log file verbosity.
LOG_FILE_MAX_FILESDESCRIPTION: Max rotated files to retain.
DESCRIPTION: Max rotated files to retain.
LOG_FILE_PATHDESCRIPTION: Log file path.
DESCRIPTION: Log file path.
LOG_LEVELinfoDESCRIPTION: Pino log level. Values: trace | debug | info | warn | error | fatal | silent
DESCRIPTION: Pino log level. Values: trace | debug | info | warn | error | fatal | silent
LYNKR_BUDGET_ENFORCERtrueDESCRIPTION: Master switch for the budget enforcer middleware. Values: true | false (set "false" to disable)
DESCRIPTION: Master switch for the budget enforcer middleware. Values: true | false (set "false" to disable)
LYNKR_CACHE_BYPASS_STRUCTUREDStructured-output requests bypass the response cache (default true).
Structured-output requests bypass the response cache (default true).
LYNKR_CASCADE_ENABLEDfalseDESCRIPTION: Enable cascading retry/escalation between tiers. Values: true | false
DESCRIPTION: Enable cascading retry/escalation between tiers. Values: true | false
LYNKR_CASCADE_VERIFYDESCRIPTION: WS6 cascade verification (EXPERIMENTAL). After a cheap-tier (SIMPLE/MEDIUM) response, run structural checks (language drift, degeneration, truncation, malformed tool calls, empty/echo) plus a coarse content score. A failing answer is DISCARDED and the request escalates up the tier ladder — a routing mistake costs seconds, not a garbage answer. Failed attempts feed the learning loop as
DESCRIPTION: WS6 cascade verification (EXPERIMENTAL). After a cheap-tier (SIMPLE/MEDIUM) response, run structural checks (language drift, degeneration, truncation, malformed tool calls, empty/echo) plus a coarse content score. A failing answer is DISCARDED and the request escalates up the tier ladder — a routing mistake costs seconds, not a garbage answer. Failed attempts feed the learning loop as
LYNKR_COST_OPTIMIZEtrueDESCRIPTION: Cost-optimized routing (downgrade tier when safe). Values: true | false
DESCRIPTION: Cost-optimized routing (downgrade tier when safe). Values: true | false
LYNKR_DECISION_HEADERSEmit X-Lynkr-Decision / X-Lynkr-Decision-Tier / X-Lynkr-Signals response headers.
Emit X-Lynkr-Decision / X-Lynkr-Decision-Tier / X-Lynkr-Signals response headers.
LYNKR_EMBEDDINGS_MAX_CHARSMax characters sent to the embedding model (default 5000; nomic-embed-text has a 2048-token context).
Max characters sent to the embedding model (default 5000; nomic-embed-text has a 2048-token context).
LYNKR_FORWARD_RESPONSE_FORMATForward the client's output_format / response_format upstream (default true).
Forward the client's output_format / response_format upstream (default true).
LYNKR_INTENT_DECAY0.7DESCRIPTION: Per-turn exponential decay applied during window scoring. weighted_score = raw_score * decay^age, where age=0 is the latest user message. Higher (~0.9) = old turns linger longer; lower (~0.5) = old turns forgotten faster. 0.7 means a complex turn from 4 messages back contributes ~24% of its raw score to the max comparison. Values: float in (0, 1] (default 0.7).
DESCRIPTION: Per-turn exponential decay applied during window scoring. weighted_score = raw_score * decay^age, where age=0 is the latest user message. Higher (~0.9) = old turns linger longer; lower (~0.5) = old turns forgotten faster. 0.7 means a complex turn from 4 messages back contributes ~24% of its raw score to the max comparison. Values: float in (0, 1] (default 0.7).
LYNKR_INTENT_WINDOW_N5DESCRIPTION: For OAuth/subscription requests, score the last N user messages instead of just the latest one. Catches "this conversation HAD a complex turn earlier" (e.g. an "audit credentials" ask 4 turns back) without inflating short follow-ups ("yes", "continue"). Combined with LYNKR_INTENT_DECAY as exponential recency weighting; the message with the highest decayed score wins. Values: positive
DESCRIPTION: For OAuth/subscription requests, score the last N user messages instead of just the latest one. Catches "this conversation HAD a complex turn earlier" (e.g. an "audit credentials" ask 4 turns back) without inflating short follow-ups ("yes", "continue"). Combined with LYNKR_INTENT_DECAY as exponential recency weighting; the message with the highest decayed score wins. Values: positive
LYNKR_KNN_CONFIDENCE_HIGH0.7DESCRIPTION: HIGH-confidence threshold above which a kNN suggestion overrides the tier-config model directly. Values in (0, 1].
DESCRIPTION: HIGH-confidence threshold above which a kNN suggestion overrides the tier-config model directly. Values in (0, 1].
LYNKR_KNN_CONFIDENCE_LOW0.4DESCRIPTION: LOW-confidence threshold. Between LOW and HIGH the suggestion is treated as "ambiguous" and — subject to the WS2 leash (only escalate when telemetry.underProvisionedPct ≥ 2%) — may bump the tier one level for safety.
DESCRIPTION: LOW-confidence threshold. Between LOW and HIGH the suggestion is treated as "ambiguous" and — subject to the WS2 leash (only escalate when telemetry.underProvisionedPct ≥ 2%) — may bump the tier one level for safety.
LYNKR_KNN_ENABLEDSkip the kNN router and its embedding call entirely.
Skip the kNN router and its embedding call entirely.
LYNKR_KNN_MIN_INDEX_SIZE100---- Learning loop (WS5) ------------------------------------------- DESCRIPTION: Minimum kNN index size before the router will use its advice. Below this size the router returns null (heuristic wins). Post-WS5 the confidence returned by query() is also damped by min(1, size/1000) so a small index advises weakly rather than not at all. Default 100 (WS5 lowered from the pre-WS5 default of 1000).
---- Learning loop (WS5) ------------------------------------------- DESCRIPTION: Minimum kNN index size before the router will use its advice. Below this size the router returns null (heuristic wins). Post-WS5 the confidence returned by query() is also damped by min(1, size/1000) so a small index advises weakly rather than not at all. Default 100 (WS5 lowered from the pre-WS5 default of 1000).
LYNKR_MAX_SESSION_TURNS200DESCRIPTION: Hard cap on total turns per session. Unset = no cap. Recommended for agentic harnesses: 200
DESCRIPTION: Hard cap on total turns per session. Unset = no cap. Recommended for agentic harnesses: 200
LYNKR_MAX_TOOL_TURNS50DESCRIPTION: Hard cap on consecutive tool-loop turns without new user text. Unset = no cap. Recommended for agentic harnesses: 50
DESCRIPTION: Hard cap on consecutive tool-loop turns without new user text. Unset = no cap. Recommended for agentic harnesses: 50
LYNKR_NATIVE_PASSTHROUGHtrueDESCRIPTION: Kill switch for native-format streaming passthrough (Anthropic client + Anthropic upstream pipes SSE bytes straight through, skipping the buffered orchestrator). Values: true | false (default true)
DESCRIPTION: Kill switch for native-format streaming passthrough (Anthropic client + Anthropic upstream pipes SSE bytes straight through, skipping the buffered orchestrator). Values: true | false (default true)
LYNKR_OAUTH_PASSTHROUGHDESCRIPTION: Forward incoming OAuth Bearer tokens straight to the upstream Anthropic endpoint. Set automatically by `lynkr wrap claude`. Values: true | false
DESCRIPTION: Forward incoming OAuth Bearer tokens straight to the upstream Anthropic endpoint. Set automatically by `lynkr wrap claude`. Values: true | false
LYNKR_OAUTH_PASSTHROUGH_URLDESCRIPTION: Upstream URL used for OAuth passthrough (defaults to AZURE_ANTHROPIC_ENDPOINT).
DESCRIPTION: Upstream URL used for OAuth passthrough (defaults to AZURE_ANTHROPIC_ENDPOINT).
LYNKR_OLLAMA_BUFFER_RESPONSEStrueDESCRIPTION: Buffer Ollama responses instead of streaming. Buffering repairs leaked <think> tags from thinking models (MiniMax) but delivers all at once. Values: true | false (default true — set false to stream)
DESCRIPTION: Buffer Ollama responses instead of streaming. Buffering repairs leaked <think> tags from thinking models (MiniMax) but delivers all at once. Values: true | false (default true — set false to stream)
LYNKR_PIN_DRIFT_MARGIN5DESCRIPTION: WS1.5 upward-drift margin. A pinned session re-decides when the latest user message's heuristic score exceeds the pinned tier's calibrated ceiling + this margin. Smaller = pins escape sooner (more re-routes, better tier fidelity); larger = stickier pins (fewer cold cache reads, risk of under-tiering). Live guidance: with tool-baseline subtraction active, trivial turns score 5-15 and m
DESCRIPTION: WS1.5 upward-drift margin. A pinned session re-decides when the latest user message's heuristic score exceeds the pinned tier's calibrated ceiling + this margin. Smaller = pins escape sooner (more re-routes, better tier fidelity); larger = stickier pins (fewer cold cache reads, risk of under-tiering). Live guidance: with tool-baseline subtraction active, trivial turns score 5-15 and m
LYNKR_PREFLIGHT_ENABLEDfalseDESCRIPTION: Run client-supplied preflight commands (cwd = workspace) and short-circuit if they all pass. Values: true | false
DESCRIPTION: Run client-supplied preflight commands (cwd = workspace) and short-circuit if they all pass. Values: true | false
LYNKR_PREFLIGHT_TIMEOUT_MS120000DESCRIPTION: Per-command timeout for preflight checks, in ms.
DESCRIPTION: Per-command timeout for preflight checks, in ms.
LYNKR_REGRET_ESTIMATORfalseDESCRIPTION: Enable the regret-estimator post-hoc routing critic. Values: true | false
DESCRIPTION: Enable the regret-estimator post-hoc routing critic. Values: true | false
LYNKR_ROUTING_CONFIGDeclarative routing (signals → decisions) — config/routing.json Path to the routing config (default: config/routing.json; see config/routing.example.json). The default file has a single `legacy` decision, so routing is unchanged until you add decisions.
Declarative routing (signals → decisions) — config/routing.json Path to the routing config (default: config/routing.json; see config/routing.example.json). The default file has a single `legacy` decision, so routing is unchanged until you add decisions.
LYNKR_SHADOW_POLICYDESCRIPTION: Shadow-mode policy name (records what an alt policy would do without applying it).
DESCRIPTION: Shadow-mode policy name (records what an alt policy would do without applying it).
LYNKR_STICKY_SESSIONStrue---- Sticky sessions (WS1) ----------------------------------------- DESCRIPTION: Cache-aware sticky session pinning. Once a session is routed, subsequent turns reuse the pinned provider/model instead of re-deciding — avoids a cold-cache re-read (~10× that turn's input cost) on every provider switch. Re-decides on compaction, guard escalation (risk/context/vision), or economic downgrade below the
---- Sticky sessions (WS1) ----------------------------------------- DESCRIPTION: Cache-aware sticky session pinning. Once a session is routed, subsequent turns reuse the pinned provider/model instead of re-deciding — avoids a cold-cache re-read (~10× that turn's input cost) on every provider switch. Re-decides on compaction, guard escalation (risk/context/vision), or economic downgrade below the
LYNKR_STICKY_TTL_MS21600000DESCRIPTION: TTL for a session pin (ms). Older pins are evicted by the 5-minute cleanup tick in src/sessions/cleanup.js. Default: 6 hours (21_600_000 ms).
DESCRIPTION: TTL for a session pin (ms). Older pins are evicted by the 5-minute cleanup tick in src/sessions/cleanup.js. Default: 6 hours (21_600_000 ms).
LYNKR_STREAM_IDLE_TIMEOUT_MS60000DESCRIPTION: Abort a passthrough stream when the upstream goes silent this long.
DESCRIPTION: Abort a passthrough stream when the upstream goes silent this long.
LYNKR_STREAM_TRANSFORMtrueDESCRIPTION: Kill switch for cross-format streaming (OpenAI upstream SSE reshaped into Anthropic events in flight instead of buffering). Values: true | false (default true)
DESCRIPTION: Kill switch for cross-format streaming (OpenAI upstream SSE reshaped into Anthropic events in flight instead of buffering). Values: true | false (default true)
LYNKR_STREAM_TRANSFORM_PROVIDERSDESCRIPTION: Providers eligible for the cross-format stream transform. Values: comma-separated (default openai,atlas,azure-openai,openrouter,databricks,lmstudio,llamacpp,moonshot)
DESCRIPTION: Providers eligible for the cross-format stream transform. Values: comma-separated (default openai,atlas,azure-openai,openrouter,databricks,lmstudio,llamacpp,moonshot)
LYNKR_SWITCH_MAX_PROMPT_TOKENSDESCRIPTION: Prompt-token cap under which an economic downgrade may override the pin. Above this the cold-cache re-read dominates the per-token savings, so the pin wins even if the fresh routing decision picks a cheaper model. Default 20 000.
DESCRIPTION: Prompt-token cap under which an economic downgrade may override the pin. Above this the cold-cache re-read dominates the per-token savings, so the pin wins even if the fresh routing decision picks a cheaper model. Default 20 000.
LYNKR_VISIBLE_ROUTINGfalseDESCRIPTION: Include `lynkr_interaction` block in every successful response showing where it routed. Values: true | false
DESCRIPTION: Include `lynkr_interaction` block in every successful response showing where it routed. Values: true | false
LYNKR_WRAP_SHOW_STATStrueDESCRIPTION: Show stats summary on exit when running `lynkr wrap claude`. Values: true | false
DESCRIPTION: Show stats summary on exit when running `lynkr wrap claude`. Values: true | false
MARKDOWN_RENDER_ANSIfalseDESCRIPTION: Render markdown to ANSI for CLIs without a markdown renderer. Leave false for Claude Code (it renders markdown itself). Values: true | false
DESCRIPTION: Render markdown to ANSI for CLIs without a markdown renderer. Leave false for Claude Code (it renders markdown itself). Values: true | false
MCP_MANIFEST_DIRS~/.claude/mcpDESCRIPTION: Comma-separated directories scanned for MCP manifests.
DESCRIPTION: Comma-separated directories scanned for MCP manifests.
MCP_SANDBOX_ADD_CAPABILITIESDESCRIPTION: Linux capabilities to add back.
DESCRIPTION: Linux capabilities to add back.
MCP_SANDBOX_ALLOW_NETWORKINGfalseDESCRIPTION: Allow network access from inside the sandbox. Values: true | false
DESCRIPTION: Allow network access from inside the sandbox. Values: true | false
MCP_SANDBOX_CONTAINER_WORKSPACE/workspaceDESCRIPTION: Workspace mount point inside the container.
DESCRIPTION: Workspace mount point inside the container.
MCP_SANDBOX_CPU_LIMIT1.0DESCRIPTION: Container CPU limit (cores).
DESCRIPTION: Container CPU limit (cores).
MCP_SANDBOX_DROP_CAPABILITIESALLDESCRIPTION: Linux capabilities to drop (comma-separated, or "ALL").
DESCRIPTION: Linux capabilities to drop (comma-separated, or "ALL").
MCP_SANDBOX_ENABLEDtrueDESCRIPTION: Master switch for the MCP sandbox. Values: true | false
DESCRIPTION: Master switch for the MCP sandbox. Values: true | false
MCP_SANDBOX_ENTRYPOINTDESCRIPTION: Override container entrypoint.
DESCRIPTION: Override container entrypoint.
MCP_SANDBOX_EXTRA_MOUNTSDESCRIPTION: Extra bind mounts (HOST:CONTAINER[:ro], comma-separated).
DESCRIPTION: Extra bind mounts (HOST:CONTAINER[:ro], comma-separated).
MCP_SANDBOX_IMAGEDESCRIPTION: Container image used when sandboxing MCP servers.
DESCRIPTION: Container image used when sandboxing MCP servers.
MCP_SANDBOX_MEMORY_LIMIT512mDESCRIPTION: Container memory limit.
DESCRIPTION: Container memory limit.
MCP_SANDBOX_MOUNT_WORKSPACEtrueDESCRIPTION: Mount the host workspace into the container. Values: true | false
DESCRIPTION: Mount the host workspace into the container. Values: true | false
MCP_SANDBOX_NETWORK_MODEnoneDESCRIPTION: Docker network mode. Values: none | bridge | host | <custom>
DESCRIPTION: Docker network mode. Values: none | bridge | host | <custom>
MCP_SANDBOX_NO_NEW_PRIVILEGEStrueDESCRIPTION: Set --security-opt no-new-privileges. Values: true | false
DESCRIPTION: Set --security-opt no-new-privileges. Values: true | false
MCP_SANDBOX_PASSTHROUGH_ENVPATH,LANG,LC_ALL,TERM,HOMEDESCRIPTION: Comma-separated env vars to forward into the sandbox.
DESCRIPTION: Comma-separated env vars to forward into the sandbox.
MCP_SANDBOX_PERMISSION_ALLOWDESCRIPTION: Comma-separated tool names always allowed.
DESCRIPTION: Comma-separated tool names always allowed.
MCP_SANDBOX_PERMISSION_DENYDESCRIPTION: Comma-separated tool names always denied.
DESCRIPTION: Comma-separated tool names always denied.
MCP_SANDBOX_PERMISSION_MODEautoDESCRIPTION: How tool permissions are decided. Values: auto | allowlist | denylist | prompt
DESCRIPTION: How tool permissions are decided. Values: auto | allowlist | denylist | prompt
MCP_SANDBOX_PIDS_LIMIT100DESCRIPTION: Max PIDs inside the container.
DESCRIPTION: Max PIDs inside the container.
MCP_SANDBOX_READ_ONLY_ROOTfalseDESCRIPTION: Mount the container root filesystem read-only. Values: true | false
DESCRIPTION: Mount the container root filesystem read-only. Values: true | false
MCP_SANDBOX_REUSE_SESSIONtrueDESCRIPTION: Reuse the same sandbox container across calls within a session. Values: true | false
DESCRIPTION: Reuse the same sandbox container across calls within a session. Values: true | false
MCP_SANDBOX_RUNTIMEdockerDESCRIPTION: Sandbox runtime. Values: docker | podman
DESCRIPTION: Sandbox runtime. Values: docker | podman
MCP_SANDBOX_TIMEOUT_MS20000DESCRIPTION: Timeout for a single MCP tool call (ms).
DESCRIPTION: Timeout for a single MCP tool call (ms).
MCP_SANDBOX_USERDESCRIPTION: Run as this user inside the container.
DESCRIPTION: Run as this user inside the container.
MCP_SERVER_MANIFESTDESCRIPTION: Single MCP servers.json manifest path.
DESCRIPTION: Single MCP servers.json manifest path.
MEMORY_DECAY_ENABLEDtrueDESCRIPTION: Apply time-based decay to memory scores. Values: true | false
DESCRIPTION: Apply time-based decay to memory scores. Values: true | false
MEMORY_DECAY_HALF_LIFE30DESCRIPTION: Half-life for memory decay (days).
DESCRIPTION: Half-life for memory decay (days).
MEMORY_DEDUP_ENABLEDtrueDESCRIPTION: Dedupe memories before injection. Values: true | false
DESCRIPTION: Dedupe memories before injection. Values: true | false
MEMORY_DEDUP_LOOKBACK5DESCRIPTION: Turns of history to scan for dedup.
DESCRIPTION: Turns of history to scan for dedup.
MEMORY_ENABLEDtrueDESCRIPTION: Master switch for the long-term Titans-inspired memory system. Values: true | false
DESCRIPTION: Master switch for the long-term Titans-inspired memory system. Values: true | false
MEMORY_EXTRACTION_ENABLEDtrueDESCRIPTION: Auto-extract memories from turns. Values: true | false
DESCRIPTION: Auto-extract memories from turns. Values: true | false
MEMORY_FORMATcompactDESCRIPTION: Memory rendering format. Values: compact | verbose | json
DESCRIPTION: Memory rendering format. Values: compact | verbose | json
MEMORY_INCLUDE_GLOBALtrueDESCRIPTION: Include cross-session global memories. Values: true | false
DESCRIPTION: Include cross-session global memories. Values: true | false
MEMORY_INJECTION_FORMATsystemDESCRIPTION: How retrieved memories are injected. Values: system | user | assistant
DESCRIPTION: How retrieved memories are injected. Values: system | user | assistant
MEMORY_MAX_AGE_DAYS90DESCRIPTION: Hard cap on memory age (days).
DESCRIPTION: Hard cap on memory age (days).
MEMORY_MAX_COUNT10000DESCRIPTION: Hard cap on total memory count.
DESCRIPTION: Hard cap on total memory count.
MEMORY_RETRIEVAL_LIMIT5DESCRIPTION: Max memories retrieved per request.
DESCRIPTION: Max memories retrieved per request.
MEMORY_SURPRISE_THRESHOLD0.3DESCRIPTION: Surprise threshold above which a turn is written to memory.
DESCRIPTION: Surprise threshold above which a turn is written to memory.
MODEL_DEFAULTDESCRIPTION: Default fallback model name when no tier/provider model is known.
DESCRIPTION: Default fallback model name when no tier/provider model is known.
MODEL_PRICE_OVERRIDESDESCRIPTION: JSON of per-model price overrides for the cost registry. Format: {"model-name":{"input":0.5,"output":1.5}}
DESCRIPTION: JSON of per-model price overrides for the cost registry. Format: {"model-name":{"input":0.5,"output":1.5}}
MODEL_PROVIDERollamaDESCRIPTION: [DEPRECATED legacy knob] primary provider for credential validation. Auto-detected when TIER_* is set. Same allowed values as FALLBACK_PROVIDER plus ollama/llamacpp/lmstudio/vertex/zai/moonshot.
DESCRIPTION: [DEPRECATED legacy knob] primary provider for credential validation. Auto-detected when TIER_* is set. Same allowed values as FALLBACK_PROVIDER plus ollama/llamacpp/lmstudio/vertex/zai/moonshot.
MOONSHOT_API_KEYMoonshot AI (Kimi) DESCRIPTION: Moonshot API key.
Moonshot AI (Kimi) DESCRIPTION: Moonshot API key.
MOONSHOT_ENDPOINThttps://api.moonshot.ai/v1/chat/completionsDESCRIPTION: Moonshot chat completions endpoint.
DESCRIPTION: Moonshot chat completions endpoint.
MOONSHOT_MODELkimi-k2-thinkingDESCRIPTION: Default Moonshot model.
DESCRIPTION: Default Moonshot model.
NODE_ENVproductionDESCRIPTION: Node runtime mode. "development" enables pino-pretty (requires install). Values: development | production | test
DESCRIPTION: Node runtime mode. "development" enables pino-pretty (requires install). Values: development | production | test
OLLAMA_EMBEDDINGS_ENDPOINThttp://localhost:11434/api/embeddingsDESCRIPTION: Embeddings endpoint URL.
DESCRIPTION: Embeddings endpoint URL.
OLLAMA_EMBEDDINGS_MODELnomic-embed-textDESCRIPTION: Embedding model (for semantic cache and Cursor @Codebase).
DESCRIPTION: Embedding model (for semantic cache and Cursor @Codebase).
OLLAMA_ENDPOINThttp://localhost:11434DESCRIPTION: Ollama server endpoint.
DESCRIPTION: Ollama server endpoint.
OLLAMA_KEEP_ALIVEDESCRIPTION: Ollama `keep_alive` parameter (e.g. "5m", "30m", "-1" for forever).
DESCRIPTION: Ollama `keep_alive` parameter (e.g. "5m", "30m", "-1" for forever).
OLLAMA_MAX_TOOLS_FOR_ROUTING3DESCRIPTION: Cap on how many tools are injected when routing through Ollama.
DESCRIPTION: Cap on how many tools are injected when routing through Ollama.
OLLAMA_MODELqwen2.5-coder:latestOllama (local models) DESCRIPTION: Default Ollama model id.
Ollama (local models) DESCRIPTION: Default Ollama model id.
OLLAMA_TIMEOUT_MS120000DESCRIPTION: Per-request timeout in milliseconds.
DESCRIPTION: Per-request timeout in milliseconds.
OPENAI_API_KEYOpenAI (direct) DESCRIPTION: OpenAI API key.
OpenAI (direct) DESCRIPTION: OpenAI API key.
OPENAI_ENDPOINThttps://api.openai.com/v1/chat/completionsDESCRIPTION: Chat completions endpoint (can point at any OpenAI-compatible host).
DESCRIPTION: Chat completions endpoint (can point at any OpenAI-compatible host).
OPENAI_MODELgpt-4oDESCRIPTION: Default OpenAI model.
DESCRIPTION: Default OpenAI model.
OPENAI_ORGANIZATIONDESCRIPTION: Optional OpenAI org id.
DESCRIPTION: Optional OpenAI org id.
OPENAI_REASONING_EFFORTFIREWORKS_THINKING_MODELS=glm-5p3|glm-5\.3|deepseek-v4p1-flash # regex: never send thinking:disabled FIREWORKS_RESPONSE_FORMAT_SCHEMA=false # Fireworks rejects json_schema with $ref; json_object by default
FIREWORKS_THINKING_MODELS=glm-5p3|glm-5\.3|deepseek-v4p1-flash # regex: never send thinking:disabled FIREWORKS_RESPONSE_FORMAT_SCHEMA=false # Fireworks rejects json_schema with $ref; json_object by default
OPENAI_RESPONSE_FORMAT_SCHEMA
OPENCLAW_MODEDESCRIPTION: OpenClaw mode — rewrites response `model` field with actual provider/model used. Values: true | false
DESCRIPTION: OpenClaw mode — rewrites response `model` field with actual provider/model used. Values: true | false
OPENROUTER_ALLOW_FALLBACKS
OPENROUTER_API_KEYOpenRouter (100+ models via single API) DESCRIPTION: OpenRouter API key.
OpenRouter (100+ models via single API) DESCRIPTION: OpenRouter API key.
OPENROUTER_EMBEDDINGS_MODELopenai/text-embedding-ada-002DESCRIPTION: Embedding model used through OpenRouter.
DESCRIPTION: Embedding model used through OpenRouter.
OPENROUTER_ENDPOINThttps://openrouter.ai/api/v1/chat/completionsDESCRIPTION: OpenRouter chat completions endpoint.
DESCRIPTION: OpenRouter chat completions endpoint.
OPENROUTER_MAX_TOKENS_MAPPer-model output cap (bounds runaway reasoning on cheap models).
Per-model output cap (bounds runaway reasoning on cheap models).
OPENROUTER_MAX_TOOLS_FOR_ROUTING15DESCRIPTION: Cap on tool count sent during routing.
DESCRIPTION: Cap on tool count sent during routing.
OPENROUTER_MODELopenai/gpt-4o-miniDESCRIPTION: Default OpenRouter model.
DESCRIPTION: Default OpenRouter model.
OPENROUTER_PROVIDER_ORDEROpenRouter: provider pin, reasoning, schema, timeouts, failover Pin provider(s) globally, or per model (';' between models, ',' between providers).
OpenRouter: provider pin, reasoning, schema, timeouts, failover Pin provider(s) globally, or per model (';' between models, ',' between providers).
OPENROUTER_PROVIDER_ORDER_MAP
OPENROUTER_REASONING_EFFORTReasoning effort (global default, or per model).
Reasoning effort (global default, or per model).
OPENROUTER_REASONING_EFFORT_MAP
OPENROUTER_RESPONSE_FORMAT_SCHEMA
OPENROUTER_SCHEMA_PROVIDERSHosts that honour strict json_schema; others receive json_object.
Hosts that honour strict json_schema; others receive json_object.
OPENROUTER_TIMEOUT_MSUpstream timeout (ms), global and per model; on timeout/5xx/404 Lynkr fails over to the next pinned provider, then lets tier-fallback climb.
Upstream timeout (ms), global and per model; on timeout/5xx/404 Lynkr fails over to the next pinned provider, then lets tier-fallback climb.
OPENROUTER_TIMEOUT_MS_MAP
ORCA_API_BASE_URLDESCRIPTION: Explicit API origin override (api.orcarouter.ai default).
DESCRIPTION: Explicit API origin override (api.orcarouter.ai default).
ORCA_AUTH_BASE_URLDESCRIPTION: Explicit auth origin override (www.orcarouter.ai default).
DESCRIPTION: Explicit auth origin override (www.orcarouter.ai default).
ORCA_BASE_URLDESCRIPTION: Shared self-hosted base URL — falls back for BOTH auth and API when the explicit overrides below are unset.
DESCRIPTION: Shared self-hosted base URL — falls back for BOTH auth and API when the explicit overrides below are unset.
ORCAROUTER_API_KEYOrcaRouter (OpenAI-compatible gateway — adaptive routing, failover, zero-markup inference, agent-tool governance) DESCRIPTION: OrcaRouter API key (sk-orca-...). You can also run `lynkr connect orcarouter` for an OAuth 2.0 + PKCE one-click login that issues a key belonging to your own account (no client secret, no redirect pre-registration). Auth requests go to www.orcarouter.ai; inference and mode
OrcaRouter (OpenAI-compatible gateway — adaptive routing, failover, zero-markup inference, agent-tool governance) DESCRIPTION: OrcaRouter API key (sk-orca-...). You can also run `lynkr connect orcarouter` for an OAuth 2.0 + PKCE one-click login that issues a key belonging to your own account (no client secret, no redirect pre-registration). Auth requests go to www.orcarouter.ai; inference and mode
ORCAROUTER_ENDPOINThttps://api.orcarouter.ai/v1/chat/completionsDESCRIPTION: Chat completions endpoint. Defaults to the API origin + /v1.
DESCRIPTION: Chat completions endpoint. Defaults to the API origin + /v1.
ORCAROUTER_MODELorcarouter/autoDESCRIPTION: Default OrcaRouter model (see the live catalog at https://api.orcarouter.ai/v1/models for your workspace's real list).
DESCRIPTION: Default OrcaRouter model (see the live catalog at https://api.orcarouter.ai/v1/models for your workspace's real list).
OVERSIZED_ERROR_LOG_DIR./logs/oversized-errorsDESCRIPTION: Where oversized-error dumps go.
DESCRIPTION: Where oversized-error dumps go.
OVERSIZED_ERROR_LOGGING_ENABLEDtrueDESCRIPTION: Persist oversized-payload errors to disk. Values: true | false
DESCRIPTION: Persist oversized-payload errors to disk. Values: true | false
OVERSIZED_ERROR_MAX_FILES100DESCRIPTION: Max retained oversized-error dump files.
DESCRIPTION: Max retained oversized-error dump files.
OVERSIZED_ERROR_THRESHOLD200DESCRIPTION: Bytes above which a payload is considered "oversized".
DESCRIPTION: Bytes above which a payload is considered "oversized".
PERF_TIMERDESCRIPTION: Print per-stage timing breakdowns to stdout. Values: true | false
DESCRIPTION: Print per-stage timing breakdowns to stdout. Values: true | false
POLICY_DISALLOWED_TOOLSDESCRIPTION: Comma-separated tool names that are never allowed.
DESCRIPTION: Comma-separated tool names that are never allowed.
POLICY_FILE_ALLOWED_PATHSDESCRIPTION: Comma-separated paths that file tools may touch (allowlist).
DESCRIPTION: Comma-separated paths that file tools may touch (allowlist).
POLICY_FILE_BLOCKED_PATHS/.env,.env,/etc/passwd,/etc/shadowDESCRIPTION: Comma-separated paths that file tools may NOT touch.
DESCRIPTION: Comma-separated paths that file tools may NOT touch.
POLICY_GIT_ALLOW_COMMITtrueDESCRIPTION: Allow `git commit`. Values: true | false
DESCRIPTION: Allow `git commit`. Values: true | false
POLICY_GIT_ALLOW_PULLtrueDESCRIPTION: Allow `git pull`. Values: true | false
DESCRIPTION: Allow `git pull`. Values: true | false
POLICY_GIT_ALLOW_PUSHfalseDESCRIPTION: Allow `git push`. Values: true | false
DESCRIPTION: Allow `git push`. Values: true | false
POLICY_GIT_AUTOSTASHDESCRIPTION: Auto-stash uncommitted changes before risky git operations. Values: true | false
DESCRIPTION: Auto-stash uncommitted changes before risky git operations. Values: true | false
POLICY_GIT_COMMIT_REGEXDESCRIPTION: Regex that commit messages must match.
DESCRIPTION: Regex that commit messages must match.
POLICY_GIT_REQUIRE_TESTSfalseDESCRIPTION: Refuse commits unless POLICY_GIT_TEST_COMMAND passes. Values: true | false
DESCRIPTION: Refuse commits unless POLICY_GIT_TEST_COMMAND passes. Values: true | false
POLICY_GIT_TEST_COMMANDDESCRIPTION: Test command run before allowing a commit (when REQUIRE_TESTS=true).
DESCRIPTION: Test command run before allowing a commit (when REQUIRE_TESTS=true).
POLICY_MAX_STEPS20DESCRIPTION: Hard cap on routing/tool-call steps per request.
DESCRIPTION: Hard cap on routing/tool-call steps per request.
POLICY_MAX_TOOL_CALLS12DESCRIPTION: Hard cap on tool calls per request.
DESCRIPTION: Hard cap on tool calls per request.
POLICY_SAFE_COMMANDS_CONFIGDESCRIPTION: JSON config for the safe-commands allowlist.
DESCRIPTION: JSON config for the safe-commands allowlist.
POLICY_SAFE_COMMANDS_ENABLEDtrueDESCRIPTION: Apply the safe-commands allowlist to bash tool calls. Values: true | false
DESCRIPTION: Apply the safe-commands allowlist to bash tool calls. Values: true | false
POLICY_TOOL_LOOP_THRESHOLD10DESCRIPTION: Force-terminate after this many same-tool calls in a row (loop guard).
DESCRIPTION: Force-terminate after this many same-tool calls in a row (loop guard).
PORT8080DESCRIPTION: HTTP port the proxy listens on.
DESCRIPTION: HTTP port the proxy listens on.
PREFER_OLLAMADESCRIPTION: [DEPRECATED] legacy preference for Ollama. Use TIER_SIMPLE=ollama:<model> instead. Values: true | false
DESCRIPTION: [DEPRECATED] legacy preference for Ollama. Use TIER_SIMPLE=ollama:<model> instead. Values: true | false
PROMPT_CACHE_ENABLEDtrueDESCRIPTION: Master switch for the in-memory prompt cache. Values: true | false
DESCRIPTION: Master switch for the in-memory prompt cache. Values: true | false
PROMPT_CACHE_MAX_ENTRIES1000DESCRIPTION: Max number of prompts to cache.
DESCRIPTION: Max number of prompts to cache.
PROMPT_CACHE_TTL_MS300000DESCRIPTION: Cache entry TTL (ms).
DESCRIPTION: Cache entry TTL (ms).
RATE_LIMIT_ENABLEDtrueDESCRIPTION: Master switch for the per-session rate limiter. Values: true | false
DESCRIPTION: Master switch for the per-session rate limiter. Values: true | false
RATE_LIMIT_KEY_BYDESCRIPTION: How requests are bucketed. Values: session | ip | both
DESCRIPTION: How requests are bucketed. Values: session | ip | both
RATE_LIMIT_MAX100DESCRIPTION: Max requests per window.
DESCRIPTION: Max requests per window.
RATE_LIMIT_WINDOW_MS60000DESCRIPTION: Sliding window length (ms).
DESCRIPTION: Sliding window length (ms).
REQUEST_JSON_LIMIT1gbDESCRIPTION: Max JSON request body size (express bodyParser units, e.g. "1gb").
DESCRIPTION: Max JSON request body size (express bodyParser units, e.g. "1gb").
RISK_TIER_ESCALATIONDisable security-keyword risk escalation (tasks ABOUT security are not risky).
Disable security-keyword risk escalation (tasks ABOUT security are not risky).
SECURITY_AUDIT_LOG_DIR./logsDESCRIPTION: Directory for the security audit log.
DESCRIPTION: Directory for the security audit log.
SECURITY_AUDIT_LOG_ENABLEDtrueDESCRIPTION: Persist security events to disk. Values: true | false
DESCRIPTION: Persist security events to disk. Values: true | false
SECURITY_BLOCK_ON_DETECTIONtrueDESCRIPTION: Block requests when the filter triggers (vs. just log). Values: true | false
DESCRIPTION: Block requests when the filter triggers (vs. just log). Values: true | false
SECURITY_CONTENT_FILTER_ENABLEDtrueDESCRIPTION: Master switch for the security content filter. Values: true | false
DESCRIPTION: Master switch for the security content filter. Values: true | false
SECURITY_PER_ENDPOINT_LIMIT1000DESCRIPTION: Per-endpoint request cap per minute.
DESCRIPTION: Per-endpoint request cap per minute.
SECURITY_PER_IP_LIMIT100DESCRIPTION: Per-IP request cap per minute.
DESCRIPTION: Per-IP request cap per minute.
SECURITY_RATE_LIMIT_ENABLEDtrueDESCRIPTION: Master switch for the security rate limiter. Values: true | false
DESCRIPTION: Master switch for the security rate limiter. Values: true | false
SEMANTIC_CACHE_ENABLEDtrueDESCRIPTION: Master switch for semantic (embedding-based) response cache. Values: true | false
DESCRIPTION: Master switch for semantic (embedding-based) response cache. Values: true | false
SEMANTIC_CACHE_MAX_ENTRIES50DESCRIPTION: Max number of cached entries.
DESCRIPTION: Max number of cached entries.
SEMANTIC_CACHE_THRESHOLD0.95DESCRIPTION: Cosine-similarity threshold for a cache hit.
DESCRIPTION: Cosine-similarity threshold for a cache hit.
SEMANTIC_CACHE_TTL_MS300000DESCRIPTION: Cache entry TTL (ms).
DESCRIPTION: Cache entry TTL (ms).
SESSION_DB_PATH./data/sessions.dbDESCRIPTION: SQLite path for session storage.
DESCRIPTION: SQLite path for session storage.
SMART_TOOL_SELECTION_MODEheuristicDESCRIPTION: Smart tool selection strategy (also feeds the routing complexity analyzer). Values: heuristic | aggressive | conservative | disabled
DESCRIPTION: Smart tool selection strategy (also feeds the routing complexity analyzer). Values: heuristic | aggressive | conservative | disabled
SMART_TOOL_SELECTION_TOKEN_BUDGETDESCRIPTION: Token budget the smart-selector tries to stay under.
DESCRIPTION: Token budget the smart-selector tries to stay under.
SUGGESTION_MODE_MODELdefaultDESCRIPTION: Suggestion-mode model override. Values: default (same as MODEL_PROVIDER) | none | <model-name>
DESCRIPTION: Suggestion-mode model override. Values: default (same as MODEL_PROVIDER) | none | <model-name>
SYSTEM_PROMPT_MODEdynamicDESCRIPTION: System prompt rendering strategy. Values: dynamic | static | minimal
DESCRIPTION: System prompt rendering strategy. Values: dynamic | static | minimal
TENCENTDB_MEMORY_CORE_PORTDESCRIPTION: Host port overrides (defaults shown).
DESCRIPTION: Host port overrides (defaults shown).
TENCENTDB_MEMORY_DOCKER_ENABLEDDESCRIPTION: Let Lynkr manage the containers. Set false if you run the stack yourself via the project's deploy scripts. Values: true | false
DESCRIPTION: Let Lynkr manage the containers. Set false if you run the stack yourself via the project's deploy scripts. Values: true | false
TENCENTDB_MEMORY_ENABLEDtrueDESCRIPTION: TencentDB-Agent-Memory sidecar (team memory hub: L0-L3 chat memory, Skills, Wiki, CodeGraph). When enabled, `lynkr start` launches the memory-core + memory-hub containers from Docker Hub (agentmemory/*) — same pattern as the Headroom sidecar. Panel UI: http://localhost:8125 Requires Docker (skips gracefully if unavailable). Containers persist across Lynkr restarts (remove with: docker
DESCRIPTION: TencentDB-Agent-Memory sidecar (team memory hub: L0-L3 chat memory, Skills, Wiki, CodeGraph). When enabled, `lynkr start` launches the memory-core + memory-hub containers from Docker Hub (agentmemory/*) — same pattern as the Headroom sidecar. Panel UI: http://localhost:8125 Requires Docker (skips gracefully if unavailable). Containers persist across Lynkr restarts (remove with: docker
TENCENTDB_MEMORY_KNOWLEDGE_PORT
TENCENTDB_MEMORY_LLM_API_KEY
TENCENTDB_MEMORY_LLM_BASE_URLDESCRIPTION: LLM endpoint the memory services use for extraction and wiki ingest. Defaults to Lynkr's own OpenAI-compatible endpoint (tier routing picks the model), so no extra API key is needed. Override to point at a provider directly.
DESCRIPTION: LLM endpoint the memory services use for extraction and wiki ingest. Defaults to Lynkr's own OpenAI-compatible endpoint (tier routing picks the model), so no extra API key is needed. Override to point at a provider directly.
TENCENTDB_MEMORY_LLM_MODEL
TENCENTDB_MEMORY_PANEL_PORT
TENCENTDB_MEMORY_PROMPT_MODEDESCRIPTION: Memory extraction style. `code` extracts changes/issues/tool usage (coding agents); `chat` extracts general conversational facts. Values: code | chat
DESCRIPTION: Memory extraction style. `code` extracts changes/issues/tool usage (coding agents); `chat` extracts general conversational facts. Values: code | chat
TIER_COMPLEXDESCRIPTION: Provider:model for complex tasks (multi-file changes, design)
DESCRIPTION: Provider:model for complex tasks (multi-file changes, design)
TIER_MEDIUMollama:qwen2.5-coder:latestDESCRIPTION: Provider:model for moderate tasks (code edits, small refactors)
DESCRIPTION: Provider:model for moderate tasks (code edits, small refactors)
TIER_REASONINGmoonshot:kimi-k2-thinkingDESCRIPTION: Provider:model for hard reasoning (algorithms, debugging)
DESCRIPTION: Provider:model for hard reasoning (algorithms, debugging)
TIER_SIMPLEollama:qwen2.5-coder:latestDESCRIPTION: Provider:model for trivial single-shot tasks (greetings, formatting)
DESCRIPTION: Provider:model for trivial single-shot tasks (greetings, formatting)
TINYFISH_API_KEYDESCRIPTION: TinyFish.ai API key (enables the WebAgent browser-automation tool).
DESCRIPTION: TinyFish.ai API key (enables the WebAgent browser-automation tool).
TINYFISH_BROWSER_PROFILEliteDESCRIPTION: Browser profile preset. Values: lite | standard | stealth
DESCRIPTION: Browser profile preset. Values: lite | standard | stealth
TINYFISH_ENDPOINThttps://agent.tinyfish.ai/v1/automation/run-sseDESCRIPTION: TinyFish automation endpoint.
DESCRIPTION: TinyFish automation endpoint.
TINYFISH_PROXY_COUNTRYUSDESCRIPTION: Proxy egress country (ISO 3166 alpha-2).
DESCRIPTION: Proxy egress country (ISO 3166 alpha-2).
TINYFISH_PROXY_ENABLEDfalseDESCRIPTION: Route the browser through a residential proxy. Values: true | false
DESCRIPTION: Route the browser through a residential proxy. Values: true | false
TINYFISH_TIMEOUT_MS120000DESCRIPTION: Per-run timeout (ms).
DESCRIPTION: Per-run timeout (ms).
TOKEN_BUDGET_ENFORCEMENTDESCRIPTION: Apply adaptive compression when a request exceeds the budget (instead of only warning). Requests are compressed to fit, not refused. Values: true | false
DESCRIPTION: Apply adaptive compression when a request exceeds the budget (instead of only warning). Requests are compressed to fit, not refused. Values: true | false
TOKEN_BUDGET_MAXDESCRIPTION: OPT-IN global request-size ceiling (tokens). LEAVE UNSET for the default behavior: the budget is derived automatically from the SERVED model's real context window (registry window × 0.85, pin-aware behind virtual model names) — a 128k model gets a 108k budget, a 1M model gets an 850k budget. Set this ONLY as a deliberate cost/latency ceiling: input tokens are re-billed every turn, so
DESCRIPTION: OPT-IN global request-size ceiling (tokens). LEAVE UNSET for the default behavior: the budget is derived automatically from the SERVED model's real context window (registry window × 0.85, pin-aware behind virtual model names) — a 128k model gets a 108k budget, a 1M model gets an 850k budget. Set this ONLY as a deliberate cost/latency ceiling: input tokens are re-billed every turn, so
TOKEN_BUDGET_WARNINGDESCRIPTION: Token budget that triggers a "you're approaching the limit" warning.
DESCRIPTION: Token budget that triggers a "you're approaching the limit" warning.
TOKEN_TRACKING_ENABLEDDESCRIPTION: Track input/output tokens per request. Values: true | false
DESCRIPTION: Track input/output tokens per request. Values: true | false
TOOL_DESCRIPTIONSminimalDESCRIPTION: How verbose tool descriptions are. Values: minimal | normal | verbose
DESCRIPTION: How verbose tool descriptions are. Values: minimal | normal | verbose
TOON_ENABLEDDESCRIPTION: Enable TOON (token-optimized object notation) encoding for large structured payloads. Values: true | false
DESCRIPTION: Enable TOON (token-optimized object notation) encoding for large structured payloads. Values: true | false
TOON_FAIL_OPENtrueDESCRIPTION: Continue without TOON on encoder failure instead of erroring. Values: true | false
DESCRIPTION: Continue without TOON on encoder failure instead of erroring. Values: true | false
TOON_LOG_STATStrueDESCRIPTION: Log per-request TOON savings stats. Values: true | false
DESCRIPTION: Log per-request TOON savings stats. Values: true | false
TOON_MIN_BYTES4096DESCRIPTION: Minimum byte size before TOON encoding kicks in.
DESCRIPTION: Minimum byte size before TOON encoding kicks in.
TYPESAFE_API_KEYDESCRIPTION: TypeSafe API key for the Jev routing judge (second opinion on intent difficulty + risk corroboration). Optional — anchor scoring carries routing alone when unset (fail-soft null, no new env vars otherwise).
DESCRIPTION: TypeSafe API key for the Jev routing judge (second opinion on intent difficulty + risk corroboration). Optional — anchor scoring carries routing alone when unset (fail-soft null, no new env vars otherwise).
VERTEX_API_KEYGoogle Vertex AI / Gemini DESCRIPTION: Vertex/Gemini API key (preferred name).
Google Vertex AI / Gemini DESCRIPTION: Vertex/Gemini API key (preferred name).
VERTEX_MODELgemini-2.0-flashDESCRIPTION: Default Gemini model.
DESCRIPTION: Default Gemini model.
WEB_FETCH_BODY_PREVIEW_MAX10000DESCRIPTION: Max bytes of page body shown in WebFetch results.
DESCRIPTION: Max bytes of page body shown in WebFetch results.
WEB_SEARCH_ALLOW_ALLtrueDESCRIPTION: Allow searching any host (overrides the allowlist). Values: true | false
DESCRIPTION: Allow searching any host (overrides the allowlist). Values: true | false
WEB_SEARCH_ALLOWED_HOSTSDESCRIPTION: Comma-separated allowlist of search hosts.
DESCRIPTION: Comma-separated allowlist of search hosts.
WEB_SEARCH_API_KEYDESCRIPTION: Bearer token for the search endpoint.
DESCRIPTION: Bearer token for the search endpoint.
WEB_SEARCH_ENDPOINThttp://localhost:8888/searchDESCRIPTION: Endpoint for the WebSearch tool (often a local SearXNG instance).
DESCRIPTION: Endpoint for the WebSearch tool (often a local SearXNG instance).
WEB_SEARCH_MAX_RETRIES2DESCRIPTION: Max retry attempts on failure.
DESCRIPTION: Max retry attempts on failure.
WEB_SEARCH_RETRY_ENABLEDtrueDESCRIPTION: Retry failed searches. Values: true | false
DESCRIPTION: Retry failed searches. Values: true | false
WEB_SEARCH_TIMEOUT_MS10000DESCRIPTION: Per-search timeout (ms).
DESCRIPTION: Per-search timeout (ms).
WORKER_OFFLOAD_THRESHOLD_BYTES10000DESCRIPTION: Payload size (bytes) above which work is offloaded to the pool.
DESCRIPTION: Payload size (bytes) above which work is offloaded to the pool.
WORKER_POOL_ENABLEDtrueDESCRIPTION: Master switch for the worker thread pool (heavy parsing, embeddings). Values: true | false
DESCRIPTION: Master switch for the worker thread pool (heavy parsing, embeddings). Values: true | false
WORKER_POOL_SIZE0DESCRIPTION: Worker pool size. 0 = auto (CPU cores - 1).
DESCRIPTION: Worker pool size. 0 = auto (CPU cores - 1).
WORKER_TASK_TIMEOUT_MS5000DESCRIPTION: Per-task timeout (ms).
DESCRIPTION: Per-task timeout (ms).
WORKSPACE_ROOT/path/to/your/workspaceDESCRIPTION: Absolute path to the workspace Lynkr operates on.
DESCRIPTION: Absolute path to the workspace Lynkr operates on.
WORKSPACE_TEST_ARGSDESCRIPTION: Extra args appended to the test command.
DESCRIPTION: Extra args appended to the test command.
WORKSPACE_TEST_COMMANDDESCRIPTION: Test command the workspace runner invokes.
DESCRIPTION: Test command the workspace runner invokes.
WORKSPACE_TEST_COVERAGE_FILEScoverage/coverage-summary.jsonDESCRIPTION: Coverage report paths the runner picks up.
DESCRIPTION: Coverage report paths the runner picks up.
WORKSPACE_TEST_PROFILESDESCRIPTION: JSON array of named test profiles.
DESCRIPTION: JSON array of named test profiles.
WORKSPACE_TEST_SANDBOXautoDESCRIPTION: How tests are sandboxed. Values: auto | docker | none
DESCRIPTION: How tests are sandboxed. Values: auto | docker | none
WORKSPACE_TEST_TIMEOUT_MS600000DESCRIPTION: Test timeout (ms).
DESCRIPTION: Test timeout (ms).
ZAI_API_KEYZ.AI (Zhipu AI) DESCRIPTION: Z.AI API key.
Z.AI (Zhipu AI) DESCRIPTION: Z.AI API key.
ZAI_ENDPOINThttps://api.z.ai/api/anthropic/v1/messagesDESCRIPTION: Z.AI Anthropic-compatible messages endpoint.
DESCRIPTION: Z.AI Anthropic-compatible messages endpoint.
ZAI_MAX_CONCURRENT2DESCRIPTION: Max concurrent in-flight Z.AI requests.
DESCRIPTION: Max concurrent in-flight Z.AI requests.
ZAI_MODELGLM-4.7DESCRIPTION: Default Z.AI model.
DESCRIPTION: Default Z.AI model.

FAQ

Can I run Lynkr completely offline?

Yes — use local providers such as Ollama or llama.cpp for chat and embeddings, which keeps all data on your machine.

What metrics does Lynkr collect?

Request rate, latency percentiles, error and token/cost metrics plus memory, CPU, connection and circuit-breaker state, exposed in Prometheus format at /metrics.

How do I deploy with Docker?

Clone the repo, copy .env.example to .env with your credentials and run docker-compose up -d, or build and run the lynkr image standalone.

Where do tools execute?

Always on the client machine; Lynkr only forwards tool_use blocks and returns results with the next request.

Does Lynkr support MCP (Model Context Protocol)?

Yes: it discovers manifests in ~/.claude/mcp, speaks JSON-RPC 2.0 to MCP servers, registers their tools dynamically and can sandbox them in Docker.

What is token optimization and how does it save costs?

Five phases — prompt caching, memory deduplication, tool-response truncation, dynamic system prompts and conversation compression — cut provider costs by a claimed 60-80%.

What is FALLBACK_PROVIDER?

A cloud provider Lynkr transparently retries against when the tier-selected provider fails; it cannot be a local provider and defaults to databricks.

What happens if I only set some TIER_* vars?

All four tier variables must be set; otherwise tier routing is disabled entirely and MODEL_PROVIDER handles every request.

How do I install Lynkr?

Globally via npm (`npm install -g lynkr` then `lynkr start`), through the Homebrew tap, or by cloning the repo and running npm start.

What's the difference between Lynkr and native Claude Code?

Native Claude Code only reaches Anthropic's cloud at full price, while Lynkr adds 12+ providers, local models, token optimization and cost tracking.

How much does Lynkr cost?

Nothing — it is free and open source under Apache 2.0; you only pay your chosen model provider.

Does Lynkr work with Cursor IDE?

Yes, through its OpenAI-compatible endpoint: set Cursor's base URL to http://localhost:8081/v1 and any non-empty API key.

What is Lynkr?

A self-hosted proxy server that lets Claude Code CLI and Cursor IDE talk to many LLM providers instead of being locked to Anthropic's API.

Rate this project

Click to rate — no account needed.

0 comments

Loading comments…

639stars

GitHub Website

Rating

Links

Lynkr documentation GitHub Discussions DeepWiki

Author	Fast-Editor
Ports	8081HTTP proxy/gateway and built-in dashboard (configurable via PORT)8787optional Headroom context-compression sidecar container
Audience	Developers
RAM	8GB+ RAM for 7-8B models, 16GB+ RAM for 14B models (local model hosting)

Requires

Node.js
Graphifyoptional
Ollamaoptional

Monetization

free

Integrations

Databricks OpenAI Moonshot AI OpenRouter Z.AI Llama.cpp DeepSeek Azure OpenAI AWS Bedrock LM Studio Fireworks AI Google Vertex AIBaidu QianfanAtlas CloudAzure AnthropicOrcaRouter

Languages

JavaScript Rust

Platforms

Docker Docker Compose Linux NPM Homebrew Build from Source

Self-host Lynkr

What is Lynkr?

Should you self-host Lynkr?

Features

How to install Lynkr

Install Lynkr with Docker

Install Lynkr with Docker Compose

Install Lynkr with Linux

Install Lynkr with NPM

Install Lynkr with Homebrew

Install Lynkr with Build from Source

How to configure Lynkr