Self-host AnythingLLM
v1.19.0All-in-one AI app for local-first agents and document chat
- Stars
- 68 289
- Port
- 3001 (HTTP server)
mintplexlabs/anythingllmWhat is AnythingLLM?
AnythingLLM is an all-in-one AI application for self-hosting that combines document chat, AI agents, and multi-user workspaces. It supports a wide range of local and cloud LLM providers (Ollama, LM Studio, OpenAI, Anthropic, and many more), multiple vector databases (LanceDB, PGVector, Chroma, Qdrant, etc.), and features like scheduled tasks, custom agent skills, model routing, and a meeting assistant. It can be deployed via Docker, desktop apps (macOS, Windows, Linux), or bare-metal. The Docker version adds multi-user support and an embeddable chat widget. Licensed MIT.
Should you self-host AnythingLLM?
AnythingLLM scores 87/100 on selfhost.directory (how scoring works) and ranks #2 of 120 in AI Assistant. It has 68 289 GitHub stars, +1 338 in the last 30 days.
It shipped 18 stable releases in the past year. The latest, v1.19.0, came out on 26 October 2026. New versions land roughly every 2 weeks.
Setup is moderate: expect some configuration beyond a single command. An official Docker image is available. It is MIT-licensed and written mostly in JavaScript.
Its closest alternative on selfhost.directory is meetily (79/100, 32 701 stars). AnythingLLM scores higher overall. See AnythingLLM vs meetily side by side.
Features
- AI Agents
- Custom agent skills (no-code builder)
- Developer API
- Document chat (RAG)
- Embeddable chat widgetDocker version only
- Image Generation
- MCP compatibility
- Meeting Assistant
- Memories
- Model Routing
- Multi-modal support
- Multiple document types
- Multi-user supportDocker version only
- Scheduled Tasks
AnythingLLM README
How to install AnythingLLM
Install AnythingLLM with Docker Compose
version: '3.8'
name: anythingllm
networks: anything-llm: driver: bridge
services: anything-llm: container_name: anythingllm build: context: ../. dockerfile: ./docker/Dockerfile args: ARG_UID: ${UID:-1000} ARG_GID: ${GID:-1000} cap_add: - SYS_ADMIN volumes: - "./.env:/app/server/.env" - "../server/storage:/app/server/storage" - "../collector/hotdir/:/app/collector/hotdir" - "../collector/outputs/:/app/collector/outputs" user: "${UID:-1000}:${GID:-1000}" ports: - "3001:3001" env_file: - .env networks: - anything-llm extra_hosts: - "host.docker.internal:host-gateway"
From docker/docker-compose.yml
AnythingLLM releases & version history
AI summary Docker now ships with a built-in local LLM experience
Built-in LLM for Docker
AnythingLLM Docker now has a default local LLM experience right out of the box. The same kind of built-in engine that ships with AnythingLLM Desktop can now run inside the container: Lemonade, the fully open-source local inference server from AMD, running your models with llama.cpp. You no longer need to run Ollama, LM Studio or any other inference server next to AnythingLLM.
Pass your GPU into the container and the engine is accelerated on whatever hardware you have. It picks the fastest backend automatically: CUDA for NVIDIA GPUs, Vulkan for AMD and Intel GPUs, or the CPU. It runs on both amd64 and arm64 hosts.
When it's turned on, Built-in becomes the default LLM provider during onboarding, with three models recommended for your hardware. You can download any model from the model table or pull any GGUF model from Hugging Face. It works everywhere a provider can be used: the system LLM, workspace and agent LLMs, the chat window model picker, and Model Router.
Tip
The built-in LLM is off by default. Turn it on by starting the container with -e ENABLE_BUILT_IN_LLM=true, and add --gpus all (or pass in your AMD or Intel GPU devices) for GPU acceleration. See Built-in LLM for Docker for the full setup for each kind of hardware.
Color palettes
Color palettes are straightforward thematic tinting you can apply to the entire UI. We plan to expand this functionality more so you can more heavily customize the UI and even layout. Learn more
What's Changed
- fix: route Bedrock India (in.) inference profiles via bedrock-runtime by @kimnamu in #6717
- Remove open-computer (moved to its own repo) by @timothycarambat in #6727
- Neutralize spreadsheet formulas in CSV chat exports by @timothycarambat in #6728
- Cap request body size to prevent unauthenticated embed DoS by @timothycarambat in #6729
- feat: show notice when response hits max output token limit (#6703) by @ManojDevarajulu in #6726
- Hold clarifying-question timeout once the user starts answering by @timothycarambat in #6730
- Move Weaviate to weaviate-client v3 by @shatfield4 in #6675
- Support .mov files for transcription by @shatfield4 in #6721
- Stop system prompt variable values from expanding other placeholders by @shatfield4 in #6731
- Apply workspace model changes to agent sessions mid-thread by @shatfield4 in #6673
- Show per-file embed progress in context overflow modal by @shatfield4 in #6639
- fix: include email body/recipients and job prompt in tool approval requests by @rtmalikian in #6732
- fix: send the full GitLab branch name when fetching repository files by @MohammadHijjawi97 in #6740
- fix: return MCP tool text without the result envelope by @Sairaj-24 in #6736
- Fix prompt compressor crash on history records skipped by convertToPromptHistory (#6737) by @ManojDevarajulu in #6738
- fix: keep current-turn attachments when the string compressor runs by @devYRPauli in #6742
- Stop sending deprecated sampling and thinking params to Gemini by @shatfield4 in #6733
- fix: count the text of messages with attachments in statsFrom by @devYRPauli in #6745
- Built-in LLM for Docker (embedded_lemonade) by @timothycarambat in #6719
- fix: truncate long citation URLs in the detail modal by @leilei3167 in #6781
- fix: default OpenAI image generation to gpt-image-2.5-flare by @rtmalikian in #6780
- fix: keep conversation history when the OpenAI compatible chat prompt is compressed by @MarMarLabs in #6777
- fix: tell the user why a password reset was refused by @MarMarLabs in #6771
- fix: report empty connector imports as failures by @MarMarLabs in #6775
- fix(telegram): preserve complete agent answers across message chunks by @jianyx1 in #6767
- fix: connect to the document collector on 127.0.0.1 by @leilei3167 in #6766
- chore: drop API key fragments from web search error messages by @u00dxk2 in #6763
- fix: crash on workspace chat settings when saved LLM provider is not in the list yet by @TMelik in #6761
- docs: remove references to removed sqlite:migrate script by @jrlprost in #6759
- fix(collector): read Excel's short date as an ISO date by @L4XB in #6758
- Add Hungarian (hu) translation by @Proxy5 in #6757
- fix: keep a CRLF file's line endings in filesystem-edit-file by @MohammadHijjawi97 in #6756
- Parse PDF and file links in the bulk link scraper by @shatfield4 in #6786
- fix(a11y): announce a reply once it has finished streaming by @opensource-joe in #6752
- fix: use the repository default branch in the GitHub connector by @Enesuygurs in #6800
- fix(paperless-ngx): keep plain-text document content as it is instead of reading it as HTML by @L4XB in #6795
- fix(confluence): keep code blocks as written instead of running them through htmlToText by @L4XB in #6794
- fix: report failed workspace embeds on API document uploads by @Tong-bit-art in #6793
- fix: regenerate and edit should still count toward daily message limit by @jrlprost in #6785
- fix: isolate watched document sync failures by @Tong-bit-art in #6804
- fix(mcp): treat unsupported ping as non-fatal so servers stay usable by @rtmalikian in #6805
- fix: fall back to default chunk overlap when it is not smaller than the chunk size by @RaphaelFakhri in #6811
- fix: use the repository default branch in the GitLab and Gitea connectors by @JoAdo07 in #6809
- fix: preserve daily message quota when deleting threads by @Max-Reisinger in #6814
- fix(ollama): keep context windows when one model show fails by @leilei3167 in #6816
- feat: instance color palettes and tabbed branding settings by @timothycarambat in #6796
- chore: bump version to 1.19.0 by @timothycarambat in #6822
- fix(mistral): support Mistral reasoning models, dedupe model list, add thinking toggle (#6821) by @timothycarambat in #6825
New Contributors
- @ManojDevarajulu made their first contribution in #6726
- @rtmalikian made their first contribution in #6732
- @Sairaj-24 made their first contribution in #6736
- @devYRPauli made their first contribution in #6742
- @jianyx1 made their first contribution in #6767
- @TMelik made their first contribution in #6761
- @jrlprost made their first contribution in #6759
- @Proxy5 made their first contribution in #6757
- @Enesuygurs made their first contribution in #6800
- @Tong-bit-art made their first contribution in #6793
- @RaphaelFakhri made their first contribution in #6811
- @JoAdo07 made their first contribution in #6809
Full Changelog: v1.18.0...v1.19.0
AI summary Fixed image handling and provider communication
What's Changed
- fix: point k8s probes at the server's /api/ping by @u00dxk2 in #6592
- fix: don't crash when removing a parsed file while an image is attached by @marmar9615-cloud in #6585
- fix: open the clicked image when attachments share a file name by @marmar9615-cloud in #6586
- fix: stop sending history attachments to providers that build prompts from raw chat history by @simonether in #6591
- fix: compare embed allowlist entries as browser origins by @marmar9615-cloud in #6584
- add dismissal comment to prevent low-effort research claims by @timothycarambat in #6608
- fix: send agent tool required params on every provider path by @wenqiw777 in #6595
- fix: honor the SDK retry count when ANYTHINGLLM_MAX_RETRIES is set by @marmar9615-cloud in #6597
- fix: save the Outlook account type so sign-in uses the right tenant by @marmar9615-cloud in #6599
- fix: show an error when a system prompt variable change fails by @marmar9615-cloud in #6604
- fix: apply TOOL_CALL_APPROVAL_TIMEOUT_MS to Telegram tool approvals by @marmar9615-cloud in #6601
- fix: convert the hourly schedule minute to UTC in half-hour time zones by @marmar9615-cloud in #6605
- fix: select entire folder after upload when picker is paginated by @YokoAC in #6610
- fix: keep numeric strings over 15 digits as text in created Excel files by @deepanshupal in #6613
- fix: restore dropped i18n placeholders in MCP and profile strings by @theluckystrike in #6614
- fix: scrape each page once when bulk links differ only by #fragment by @MohammadHijjawi97 in #6619
- fix: attach requested files to Outlook replies and reply drafts by @marmar9615-cloud in #6623
- fix: return 404 for an unknown workspace on manage-users and update-pin by @marmar9615-cloud in #6625
- fix: load the next page of documents from the server's offset in the picker by @marmar9615-cloud in #6628
- fix: show an error when embedding attached files into the workspace fails by @marmar9615-cloud in #6629
- fix: restore dropped i18n values in model router and model picker strings by @theluckystrike in #6631
- fix: preserve referenced MBOX chat attachments during orphan cleanup by @dakjdakd in #6633
- fix: isolate chat attachment state and uploads by conversation by @dakjdakd in #6635
- fix: finalize stopped scheduled job runs on Windows by @dakjdakd in #6637
- fix: don't treat YouTube channel URLs as video links in the collector by @theluckystrike in #6644
- fix: apply workspace similarity threshold to agent rag-memory search by @marmar9615-cloud in #6648
- fix: write bulk scraped pages that share a file name to separate files by @marmar9615-cloud in #6650
- fix: surface Confluence load failures instead of reporting no pages by @marmar9615-cloud in #6652
- fix: restore document embedding for AstraDB by @marmar9615-cloud in #6654
- fix: stop scheduled jobs after switching to multi-user mode by @marmar9615-cloud in #6656
- fix(server): keep $$ and $& intact when expanding prompt variables by @rome-xi in #6676
- fix(collector): stop treating binary types with no converter as text by @moneebullah25 in #6670
- fix: use response.send instead of response.text in repo provider check by @Saadanjum0 in #6677
- fix: use a realistic fallback divisor for collector token estimates by @moneebullah25 in #6672
- fix: return after resolve in wipeCollectorStorage guards by @moneebullah25 in #6671
- Fix raw think blocks in scheduled job responses after tool calls by @shatfield4 in #6663
- fix: keep unsent prompt text when attaching a file creates a new thread by @timothycarambat in #6704
- fix: retry watched documents when re-embedding fails by @marmar9615-cloud in #6701
- fix: report a failed first Paperless-ngx request instead of no documents found by @marmar9615-cloud in #6699
- fix: keep the configured Paperless-ngx URL when following page links by @marmar9615-cloud in #6698
- fix: pass transcription and OCR settings to linked files by @marmar9615-cloud in #6697
- fix: give the same login error for an unknown username and a wrong password by @marmar9615-cloud in #6693
- fix: render nested and repeated lists correctly in created Word documents by @wenqiw777 in #6688
- fix: request serial tool use for Anthropic agent providers by @dakjdakd in #6708
- fix: propagate PGVector transaction failures without losing the original error by @dakjdakd in #6706
- fix: cancel queued scheduled-job runs when a job is disabled or resynced by @dakjdakd in #6710
- v1.18.0 tags by @timothycarambat in #6711
New Contributors
- @u00dxk2 made their first contribution in #6592
- @YokoAC made their first contribution in #6610
- @deepanshupal made their first contribution in #6613
- @theluckystrike made their first contribution in #6614
- @dakjdakd made their first contribution in #6633
- @rome-xi made their first contribution in #6676
- @moneebullah25 made their first contribution in #6670
- @Saadanjum0 made their first contribution in #6677
Full Changelog: v1.17.0...v1.18.0
AI summary Fixed spreadsheet date handling and improved model token management
What's Changed
- fix(collector): read spreadsheet dates as dates, not serial numbers by @L4XB in #6478
- Drop temp param for Anthropic models by @shatfield4 in #6472
- Allow Generic OpenAI max tokens of 0 to omit max_tokens from request by @shatfield4 in #6470
- fix: support SearchApi engines in agent web browsing by @evilive3000 in #6465
- fix: ignore malformed agent WebSocket frames instead of crashing the server by @timothycarambat in #6493
- fix(collector): read an HTML-only mbox message instead of skipping it by @L4XB in #6484
- fix(collector): read text uploads in the encoding they were written in by @L4XB in #6486
- fix(weaviate, astra): score chunks by cosine similarity like the other providers by @L4XB in #6488
- fix(bedrock): support OpenAI models on AWS by @hahahahahayesyeseys in #6473
- fix: fetch Kokoro voices from the same base path speech uses by @marmar9615-cloud in #6459
- feat(togetherai): allow configuring max_tokens to prevent response truncation by @imanecdoche in #6463
- fix: reset vectors and warn when the Azure embedding deployment changes by @marmar9615-cloud in #6461
- fix: keep manually set env settings when a settings save rewrites .env by @marmar9615-cloud in #6457
- fix: use the workspace refusal text when an embed has no documents by @marmar9615-cloud in #6454
- fix: export only the chats the chat window shows by @marmar9615-cloud in #6452
- fix: keep every sheet of a multi-sheet XLSX upload by @marmar9615-cloud in #6450
- Include generated images in chat exports by @shatfield4 in #6482
- feat: add Firecrawl agent web-search provider by @JuampiHernandez in #6480
- Autosave workspace settings on edit by @shatfield4 in #6343
- Add keyboard navigation and autofocus to provider search menus by @shatfield4 in #6483
- Read Anthropic replies from text blocks, not the first block by @timothycarambat in #6496
- fix: keep Telegram link URLs intact when they contain markdown characters by @MohammadHijjawi97 in #6524
- feat: add Gemini as an image generation provider by @petre in #6520
- fix(ui): use dynamic viewport height for mobile sidebar drawers by @albiea in #6518
- fix: keep compressed RAG context when the prompt or a chunk says "Context:" by @breken-ai in #6503
- fix: skip emoji when reading a message aloud by @SulimanAbdulrazzaq in #6500
- fix: clear the workspace router when switching back to the system default by @marmar9615-cloud in #6506
- fix: skip saving an embed chat when the reply came back empty by @marmar9615-cloud in #6508
- fix: let users and admins clear a bio by @marmar9615-cloud in #6510
- fix: resolve relative links against the page they appear on when crawling deeper by @marmar9615-cloud in #6512
- fix: keep each mbox message's own title when an upload supplies one by @marmar9615-cloud in #6514
- Add llmman image generation provider by @shatfield4 in #6497
- fix: let Milvus and Zilliz embed into workspaces whose slug starts with a digit by @wenqiw777 in #6527
- fix: filter Google Calendar skill in multi-user mode and apply its disabled sub-skills by @timothycarambat in #6539
- fix: keep Confluence code blocks whose macro does not list ac:name first by @marmar9615-cloud in #6530
- fix: embed every document an API upload returns into addToWorkspaces by @marmar9615-cloud in #6532
- fix: keep IPs, versions and decimal commas as text in created Excel files by @marmar9615-cloud in #6534
- fix: only run Telegram bot commands at the start of a message by @marmar9615-cloud in #6536
- fix: send the Keep Alive setting on agent requests to Ollama and llmman by @marmar9615-cloud in #6538
- fix(mobile): reject unknown threadSlug on reset-chat and stream-chat by @timothycarambat in #6540
- fix: render c++, ~~~ and nested fenced code as code in Telegram replies by @huiq777 in #6547
- fix: decode DuckDuckGo result links once and strip HTML from titles by @simonether in #6544
- Clean up TTS text for cloud voices and Telegram voice replies by @shatfield4 in #6552
- fix: use each workspace's prompt and id in JSONL and Alpaca chat exports by @marmar9615-cloud in #6551
- Rotate empty chat greetings by time of day by @timothycarambat in #6567
- fix: mask the saved LiteLLM API key and save changes to it by @marmar9615-cloud in #6556
- fix: shift the weekday and day of month of scheduled jobs across UTC midnight by @marmar9615-cloud in #6558
- fix: load Fireworks models and Kokoro voices with the saved API key by @marmar9615-cloud in #6557
- fix: give each new agent flow block its own id by @marmar9615-cloud in #6560
- fix: save turning password protection off so it survives a restart by @marmar9615-cloud in #6562
- fix: save links whose URL path cannot be percent-decoded by @MohammadHijjawi97 in #6565
- fix: use the embed's message history limit when the model router is on by @MohammadHijjawi97 in #6566
- fix: stop filesystem-edit-file expanding $ patterns in newText by @wenqiw777 in #6569
- fix: parse developer API document attachments for agent chats by @timothycarambat in #6573
- fix: keep uploaded images when switching from chat to @agent mid-session by @timothycarambat in #6572
- Apply workspace temperature to agents and omit the parameter when unset by @shatfield4 in #6285
- Reasoning effort controls by @shatfield4 in #6269
New Contributors
- @evilive3000 made their first contribution in #6465
- @hahahahahayesyeseys made their first contribution in #6473
- @imanecdoche made their first contribution in #6463
- @JuampiHernandez made their first contribution in #6480
- @MohammadHijjawi97 made their first contribution in #6524
- @albiea made their first contribution in #6518
- @breken-ai made their first contribution in #6503
- @SulimanAbdulrazzaq made their first contribution in #6500
- @wenqiw777 made their first contribution in #6527
- @huiq777 made their first contribution in #6547
- @simonether made their first contribution in #6544
Full Changelog: v1.16.2...v1.17.0
New Features
- LLMman - OSS is now available as a local LLM provider, replacing Docker Model Runner
- Google Vertex AI is now available as an LLM provider
- New
generate-imageagent skill for creating images from chat - You.com is now the default web search provider for richer web-searches with Zero Data Retention
- Agent flow variables now support optional arguments
- Hindi, Bulgarian, and Swedish language support added
- GenieX on Windows on ARM now lists models found in your local model directory and lets you pull models directly from Hugging Face in the UI
- Click the version tag in the top of the window to jump to your logs and storage directory more easily
Improvements
- Web browsing is now enabled by default for agents
- Local providers now report model loading status in chat so you know why the first response is slow
- Ollama runtime bumped to 0.33.3 with an updated bundled model list
- Removed the borders around the main content area for a cleaner desktop layout
- Redesigned the "new feature" modal layout
- Dates in tables are now shown in a human-readable format
- Scraped web pages now preserve tables, ordered lists, and code blocks instead of flattening them
- PDF pages are now joined with newlines so words don't fuse across page boundaries
- Mistral embeddings are now batched, making large document embedding significantly faster
- Document search now matches your terms literally instead of as a regular expression
- Watched documents now clear stale vectors on sync
- AnySearch as a web-search provider
- KoboldCPP provider overhauled and modernized to match the rest of our providers
- Keenable added as a web search provider
- OpenRouter service tier selection for controlling routing cost and priority
- LocalAI image editing via reference images
All Changes
- bump to patched @mintplex-labs/mdpdf pkg by @timothycarambat in #6216
- Identify AnythingLLM Perplexity requests by @qirh in #6206
- Add
X-Client-Infoto You.com attribution by @timothycarambat in #6221 - Google Vertex LLM Provider support by @timothycarambat in #6218
- fix: stop fillSourceWindow reversing the caller's chat history array by @marmar9615-cloud in #6231
- fix: keep workspace prompt and temperature when an embed sends no override by @marmar9615-cloud in #6232
- pass include status on sessionID chat getter by @timothycarambat in #6237
- fix: stop $ sequences corrupting restored Telegram code blocks by @marmar9615-cloud in #6235
- Add Swedish frontend locale by @yeager in #6227
- fix: keep thread and user on the query-mode refusal row by @marmar9615-cloud in #6236
- feat: replace the Docker Model Runner provider with llmman by @ericcurtin in #6226
- fix: return the API Call block's response body instead of a parse message by @marmar9615-cloud in #6240
- fix: report a failed password update instead of always answering success by @marmar9615-cloud in #6245
- fix: mobile reset-chat clears the whole workspace instead of the named thread by @marmar9615-cloud in #6241
- fix: Only block explicit matches for built-in slash commands by @marmar9615-cloud in #6243
- 6246 pw restrictions by @timothycarambat in #6250
- Pretty dates on tables by @timothycarambat in #6251
- Implement LocalAI image editing via ref_images by @TheGoodMorty in #6222
- fix: score Lance chunks at or past orthogonal as 0, not a perfect match by @marmar9615-cloud in #6253
- Migrate threads on MuM migration by @timothycarambat in #6259
- fix: stop citation cleanup mangling scraped code and numbered links by @marmar9615-cloud in #6252
- fix: scope website-depth crawling by origin and path segment by @marmar9615-cloud in #6258
- docs: fix dead links in README by @1t1sCooL in #6193
- Add variable categories to agent flow variables by @shatfield4 in #6249
- Add generate-image agent skill by @shatfield4 in #6140
- fix: four server-side bugs (privatemode provider, prompt-history delete, success flag, event log) by @simpleqt in #6165
- feat: support batch URL uploads by @Max-Reisinger in #6175
- fix: restrict workspace prompt-history read to admin/manager by @IdanMit in #6137
- fix: match document search terms literally instead of as regular expressions by @marmar9615-cloud in #6268
- fix: store DrupalWiki page attachments instead of skipping every one by @marmar9615-cloud in #6266
- fix: separate PDF page boundaries instead of fusing the adjoining words by @marmar9615-cloud in #6264
- fix(i18n): correct six Italian UI translations by @Carleoinserra in #6317
- feat(i18n): add Hindi (hi) language support by @tejasashinde in #6319
- fix: clear stale document_vectors rows during watched document sync by @timothycarambat in #6323
- fix path resolution double slash on fastCRW by @timothycarambat in #6324
- fix(collector): treat a blank published value as absent, not as the epoch by @marmar9615-cloud in #6275
- fix: keep chat-history attachments for TogetherAI, SambaNova, Z.ai and Cerebras by @marmar9615-cloud in #6296
- fix: keep the GitLab repo scheme so http self-hosted instances can re-sync by @marmar9615-cloud in #6294
- fix: pin @qdrant/js-client-rest to 1.18.0 so lockfile-less builds keep search() by @timothycarambat in #6326
- fix: honor captureAs html on the scraper's fetch fallback by @marmar9615-cloud in #6292
- fix: preserve Unicode characters in imported skill names by @dawNotPoi in #6286
- fix: require sql_query, not table_name by @timothycarambat in #6329
- fix: update workspace chat lastUpdatedAt by @AbarnaaSree in #6288
- fix(collector): keep upload-link metadata when the link resolves to a file by @marmar9615-cloud in #6277
- fix(generic-openai): honor max tokens in Tooled/UnTooled agent streams by @Jkand in #6331
- Make web search use You.com as the default provider by @shatfield4 in #6280
- Enable web-browsing by default by @timothycarambat in #6335
- Add Keenable web search provider by @timothycarambat in #6336
- Add
TOOL_CALL_APPROVAL_TIMEOUT_MSto admin env config by @timothycarambat in #6339 - Add Bulgarian (bg) translation by @axander44 in #6344
- docs: add Easypanel to self-hosting deploy table by @Ahson-Shaikh in #6360
- patch GHSA-8wh8-3v68-whwj by @timothycarambat in #6361
- fix: accept the documented allowlist_domains array in the embed API by @marmar9615-cloud in #6358
- fix: drop a GitLab access token the instance rejects instead of reusing it by @marmar9615-cloud in #6356
- fix: set Simple SSO temporary auth token expiry to 10 minutes by @timothycarambat in #6362
- fix: paperless-ngx resync returned empty for every existing document by @auyua9 in #6348
- Add
deepseek-flashto IncludeReasoning array due to deprecation/upgrade of alias by @timothycarambat in #6379 - fix(mistral): batch embedding requests and declare the engine's limits by @L4XB in #6378
- fix: stop section and cite cleanup mangling scraped code and links by @marmar9615-cloud in #6371
- fix: keep the Confluence page scheme so http self-hosted instances re-sync by @marmar9615-cloud in #6372
- fix: linked mbox scraping returned only the first message by @marmar9615-cloud in #6374
- fix: decode GitHub file content as UTF-8 when re-syncing a watched document by @marmar9615-cloud in #6373
- fix: llmman model discovery ignored the token entered in settings by @marmar9615-cloud in #6376
- KoboldCPP overhaul by @L4XB in #6380
- fix(xlsx): escape spreadsheet cells the way CSV requires by @L4XB in #6381
- fix: log chunk header exceeding chunk size by @L4XB in #6386
- fix: convert a scraped table into a table, not into loose paragraphs by @L4XB in #6387
- Add OpenRouter service tier selection by @shatfield4 in #6337
- Bump embed submodule by @shatfield4 in #6389
- Report model loading status for local agent providers by @shatfield4 in #6392
- fix: give a merged table cell the columns it covers by @L4XB in #6388
- Fix JSON body corruption when agent flow variables contain quotes by @shatfield4 in #6402
- fix(localai): pass API key when fetching generated image URL by @leilei3167 in #6425
- fix: read a system message's content parts in the OpenAI-compatible chat endpoint by @marmar9615-cloud in #6424
- fix: return 404 from thread stream-chat for a missing workspace or thread by @marmar9615-cloud in #6423
- fix: accept a GitLab clone URL, trailing slash or view URL the way the siblings do by @marmar9615-cloud in #6420
- fix: return the TogetherAI context window as a number instead of a promise by @marmar9615-cloud in #6418
- fix: carry the Confluence personal access token through the re-sync payload by @marmar9615-cloud in #6414
- Resolve GHSA-3f4c-xj8r-h4cr by @timothycarambat in #6427
- fix: stop the open-computer scraper deleting brackets from code and labels from links by @shoemoney in #6408
- fix: add static promptWindowLimit to MoonshotAI and PPIO by @timothycarambat in #6438
- fix: reject an empty message list with the existing 400 in the OpenAI-compatible chat endpoint by @marmar9615-cloud in #6431
- fix: let Groq's current vision model receive image attachments by @marmar9615-cloud in #6429
- fix: re-sync watched GitLab issues and wiki pages by @marmar9615-cloud in #6433
- feat: add AnySearch agent web-search provider by @tbontb-iaq in #6439
New Contributors
- @qirh made their first contribution in #6206
- @yeager made their first contribution in #6227
- @ericcurtin made their first contribution in #6226
- @TheGoodMorty made their first contribution in #6222
- @1t1sCooL made their first contribution in #6193
- @Max-Reisinger made their first contribution in #6175
- @IdanMit made their first contribution in #6137
- @Carleoinserra made their first contribution in #6317
- @tejasashinde made their first contribution in #6319
- @dawNotPoi made their first contribution in #6286
- @AbarnaaSree made their first contribution in #6288
- @Jkand made their first contribution in #6331
- @axander44 made their first contribution in #6344
- @Ahson-Shaikh made their first contribution in #6360
- @auyua9 made their first contribution in #6348
- @L4XB made their first contribution in #6378
- @leilei3167 made their first contribution in #6425
- @shoemoney made their first contribution in #6408
- @tbontb-iaq made their first contribution in #6439
Full Changelog: v1.16.1...v1.16.2
Activity and Chain of Thought Overhaul
We overhauled our UI for long-horizon agentic tasks that show multiple thinking and tool calls to now roll up into a single clean, collapsible UI element.
This gives you more screen real estate for the actual response while still keeping the important call stack and thoughts the model produces in an easy-to-read format.
Navigation Warning
Recently, we added true abort signals across every supported LLM so that when you cancel inference or leave the page, we stop inference to save you on compute costs.
However, this had the unintended side effect of canceling inference when navigating away from a chat, preventing background completion. Now we warn you before aborting the response.
Roadmap item: We intend to make this user configurable so you can customize this behavior and allow background inference for those with more powerful or high-concurrency setups where multiple parallel inference sessions are feasible.
Foundry Local Embedded (Desktop Only)
FoundryLocal is an inference engine built by Microsoft that is pre-installed and available on Windows 11 (version 24H2, build 26100 or later).
This uses WinML to load models for your CPU, GPU, or NPU across any Windows hardware configuration. It's a great option for hardware configurations we don't have explicit support for. Keep in mind that model availability may be limited as FoundryLocal is still rolling out. Vision models are not available currently.
We partner with Microsoft for this, if you have bugs or issues we can help forward them to the FoundryLocal team
GenieX for Snapdragon (Desktop Only)
Note
This replaces our previous Snapdragon NPU engine. Any previous models from the old engine are deleted on update, and you'll need to re-download these optimized models. You should see a significant performance increase for NPU models and regular GGUFs.
We highly recommend using only GGUF models for this engine. Qualcomm NPU-only models have 4K context windows and will result in a poor agent experience.
GenieX is an open-source inference engine built by Qualcomm (previously NexaAI) that efficiently runs GGUFs and Qualcomm AI Hub models on NPU for X and X2 Elite devices.
AnythingLLM offers fully managed and built-in support by selecting the "AnythingLLM GenieX" provider in the dropdown. Both GGUFs and NPU-only models can leverage the NPU, giving you the power efficiency and intelligence of both.
We partner with Qualcomm for this, if you have bugs or issues we can help forward them to the GenieX team or you can make an issue on their GitHub. s/o @zhiyuan8, @alexchen4ai & @alanzhuly
New Features
- Native Foundry Local SDK on Windows (x64/ARM64) - the Foundry provider is now fully self-contained and no longer requires a separate Foundry Local install, with GPU/NPU-optimized model variants exposed in the catalog
- GenieX runtime support for Windows ARM64 devices
- AWS Bedrock cross-region inference profile support - geo-prefixed profile IDs (
us.,eu.,global.) now appear in the model dropdown and route correctly, plus manual region entry for regions outside the preset list (GovCloud, specialized partitions) - LocalAI added as an image generation provider, listing only image-capable models from your server
- LocalAI context window auto-detection - context windows are read per-model from your server's config, with the manual setting now an optional override
- Silent/headless install and uninstall flags for the Windows installer - see the Windows installation docs
Improvements
- Chain of thought and agent activity are now a single collapsible component - one line when collapsed, one step per thought with a connected rail when expanded, and the live thought shown in the header while streaming
- The new chain of thought rollup is now mirrored in the Assistant Panel
- Leaving a thread or navigating away mid-inference now warns you before the response is lost, instead of silently killing the generation
- Foundry Local no longer prompts for setup when using the built-in runtime, and auto-loads models by capability detection
- Onboarding light mode styling and a redesigned LLM selection UI
- Beacon now renders thinking blocks and markdown formatting correctly while streaming, with collapsible thinking and a fix for model detail text overflow
- Windows installer now includes a sidebar panel image
- Prompt input drafts now persist between navigation
Bug Fixes
- Fixed a Gemini crash caused by a content-header length mismatch on unpinned
undiciversions - now compatible with both v6 and v7 - Fixed a crash when aborting an Anthropic (or Bedrock Anthropic) response mid-stream in
chatmode and then sending a follow-up - Fixed a memory leak from leaked abort-stream listeners, which also caused pausing a response to wipe the entire chat
- Fixed a crash when creating a new workspace
- Fixed missing setup CTAs on modals introduced by the new uniform modal component - Scheduled Job skill setup on both the category and tool views, Community Hub connection key, and the experimental features reject button
- Fixed the intermediate loader being too tall while waiting for an agent capability response
- Fixed the stray border around the delete button in Workspace Chats
- Fixed reasoning getting stuck on "thinking" when a model never closes its think tag
- Fixed agent websocket sends firing while the connection was still opening
All Changes
- fix: OCR 1-bit scanned PDFs instead of dropping pages by @tzlwn1 in #6128
- New AnythingLLM wordmark + fix logo theme resolution by @shatfield4 in #6138
- fix: persist API @agent chats into thread history by @LHMQ878 in #6125
- fix: enforce workspace memory limit during extraction by @uczltw6 in #6149
- fix(a11y): announce that a reply is generating, and that one failed by @opensource-joe in #6135
- Feat/cost tracking by @timothycarambat in #6150
- Fix agent backend crash from oversized tool reranker query by @shatfield4 in #6152
- Uniform modal component by @shatfield4 in #5741
- chore: drop two package scripts whose targets are gone by @marmar9615-cloud in #6155
- docs(gcp): describe what the deployment actually creates by @marmar9615-cloud in #6154
- fix(docker): check the port the server is actually listening on by @marmar9615-cloud in #6159
- docs: fix remaining broken relative links and dead anchors by @simpleqt in #6168
- Route Bedrock cross-region inference profiles via bedrock-runtime by @timothycarambat in #6181
- Patch crash with Anthropic mid-stream abort/reply by @timothycarambat in #6186
- LocalAI context window auto-detection by @shatfield4 in #6177
- Add LocalAI as an image generation provider by @shatfield4 in #6176
- Revamp chain of thought UI component by @shatfield4 in #6182
- Prevent unload if under inference workload by @timothycarambat in #6190
- fix: agent WSS send while connecting + prompt input draft persistence by @timothycarambat in #6191
- 1.16.1 tags by @timothycarambat in #6194
New Contributors
- @tzlwn1 made their first contribution in #6128
- @LHMQ878 made their first contribution in #6125
- @uczltw6 made their first contribution in #6149
- @opensource-joe made their first contribution in #6135
- @marmar9615-cloud made their first contribution in #6155
- @simpleqt made their first contribution in #6168
Full Changelog: v1.16.0...v1.16.1
@ Every self-hosted release, weekly
One Monday email with the week's stable releases across the directory.
>_ Models for AnythingLLM
Open-weight models you can plug into AnythingLLM, most popular first.
More in AI Assistant
AppFlowyA self-hosted AI workspace that lets teams run projects, wikis, and knowledge bases on their own infrastructure with private AI and complete data ownership.
arkonArkon is a self-hosted enterprise AI knowledge hub compiling docs into a wiki served via MCP to LLMs. Features department-scoped access, RBAC, and traceable MRP pipeline.
Ars-noteArs-note is a self-hosted, open-source Markdown workspace for game dev teams with local vault, knowledge graph, AI planning, and real-time sync.
Atomic ChatSelf-hosted desktop/mobile app for running LLMs locally with OpenAI-compatible API. 1000+ models, multiple engines, MCP, agent launching.
AuralSelf-hosted open-source AI interview platform conducting structured voice, chat, and video interviews with real-time adaptive follow-ups and AI-generated reports.
AutoMateSelf-hosted AI assistant with 30+ tools, stealth browser, and multi-channel support (Web UI, Discord, CLI). Connects to any OpenAI-compatible LLM.
AVA CoreOpen-source, self-hosted AI voice agent for Asterisk/FreePBX telephony with realtime providers or modular STT-LLM-TTS pipelines, managed from a web Admin UI.
BankMCP™BankMCP™ is a self-hosted, open-source MCP server connecting AI assistants to 2,700+ European banks via Enable Banking for read-only account access.
bbarit-ossA terminal AI coding agent distributed as one Rust binary with 15+ LLM providers and 1,000+ models — a self-hosted alternative to Claude Code.
BelullamaSelf-hosted AI app bundling Ollama, Open WebUI, and Automatic1111 for local LLMs and image generation with one-command Docker install.
BestieBestie is a self-hosted, open-source AI companion runtime with SQLite memory, Telegram/Zalo channels, cron, permission-gated tools, and a local Web UI.
Big-AGISelf-hosted, open-source AI workspace for experts: run 2 to 24 models in parallel with Beam, personas, image generation, voice and BYO API keys via Docker.

0 comments