Self-host muxi
1.xOpen-source AI application server for agent formations
- Stars
- 187
- CPU
- 2 cores
- RAM
- 2 GB
- Storage
- 10 GB
- Ports
- 7890 (MUXI Server HTTP API), 8001 (Formation runtime (first formation)), 9000 (Formation runtime range end)
ghcr.io/muxi-ai/serverWhat is muxi?
MUXI is an open-source AI application server designed to deploy and operate complete agent systems in production. Instead of writing orchestration code in frameworks, you define agents, tools, memory, knowledge, triggers, and policies as declarative YAML formations. The MUXI Server orchestrates formations, the Runtime executes them, and the CLI handles deployment. Key capabilities include: intelligent orchestration via the Overlord that routes and decomposes requests; layered, event-sourced memory with user/group/formation scopes; group-based RBAC with inheritance down to individual tools; multi-tenant isolation so one formation serves many users; proactive formations with heartbeats and scheduled tasks; MCP support in both directions (consuming and exposing tools); and 12 official SDKs across major languages. MUXI runs as a single binary on Linux, macOS, and Windows, with Docker images available. The Server and Runtime are licensed under Elastic License 2.0 (source-available), while CLI, SDKs, and schemas are Apache 2.0. It requires a minimum of 2 CPU cores, 2 GB RAM, and 10 GB disk. Long-term memory uses PostgreSQL for multi-tenancy or SQLite for single-user deployments.
Should you self-host muxi?
muxi scores 48/100 on selfhost.directory (how scoring works) and ranks #110 of 209 in AI Agents. It has 187 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 moderate: expect some configuration beyond a single command. An official Docker image is available. It is Other-licensed.
Its closest alternative on selfhost.directory is Orca (91/100, 108 642 stars). It scores higher overall, so compare the two before you pick. See muxi vs Orca side by side.
Features
- 12 official SDKs
- A2A collaboration
- Built-in Observability
- Coding-agent delegation
- Declarative agent formations (YAML)
- Formation self-tuning
- Group-based RBAC
- Intelligent orchestration (Overlord)
- Layered event-sourced memory
- MCP support (client and server)
- Multi-tenant isolation
- Persistent artifact memory
- Proactive formations
- Production resilience
- Reasoning RAG
muxi README
FAQ
Are there SDKs?
Yes, MUXI ships SDKs for Python, TypeScript, Go, Java, C#, Ruby, PHP, Rust, Kotlin, Dart, Swift, and R.
Can users bring their own API keys?
Yes, per-user credentials allow each user to store their own API keys for tools. The agent uses the user's key for their requests, while other users' requests use other keys or the formation default.
How are secrets and API keys handled?
Encrypted at rest. API keys are never stored in plain text. Per-user credentials are supported so each user can store their own keys for tools.
Can external systems trigger agents?
Yes, via HTTP webhooks, cron-style scheduled tasks, and natural language scheduling (e.g., "remind me tomorrow").
Can I stream responses in real-time?
Yes, MUXI supports SSE (Server-Sent Events) for streaming text, progress, planning, tool activity, and typed UI widgets. Non-streaming requests receive the same structured output via response envelopes.
What if I need human approval for certain actions?
Use human-in-the-loop (HITL). Define which actions require approval with conditions (e.g., amount > 1000) and notification targets. The agent pauses, notifies the approver, and continues only after approval.
Can I schedule recurring tasks?
Yes, scheduled tasks support cron-style or natural language scheduling (e.g., "every monday at 6am").
How do I reduce LLM costs?
Built-in features include semantic LLM caching (70%+ typical savings), user synopsis caching (80%+ token reduction), efficient tool loading (schemas indexed once), and model mixing (cheaper models for simple tasks).
Is it reliable enough for production?
Yes, enterprise resilience patterns are baked in: circuit breakers, retries with exponential backoff, graceful degradation, LLM response caching, and auto-restart for crashed formations.
Can I see what users are asking about?
Yes, automatic topic tagging extracts 1-5 semantic topic tags from every request during analysis (no extra LLM calls). Use for dashboards, filtering, trend analysis, and alerting.
How do I monitor my agents?
Built-in observability emits structured events for LLM calls, tool usage, workflow steps, and errors. Export to Datadog, Elastic, Splunk, OpenTelemetry, or webhooks without sidecars.
How do I deploy to production?
One command: muxi deploy. MUXI is a single binary with no dependencies (Docker optional). Supports GitOps workflows for PR-based deployments.
Can users recall past conversations?
Yes, cross-session memory allows users to reference past chats without specifying which session. Working memory spans sessions.
Can the same user chat on Slack AND web?
Yes, multi-identity lets you link multiple identifiers (email, Slack ID, internal ID) to one user so they share the same memory across platforms.
Can multiple users share one deployment?
Yes, multi-tenancy is built in. Each user gets isolated memory, separate conversation history, and their own API credentials for tools. One server can serve many users with complete isolation.
More in Application Management
Nginx-Reverse-EmbySelf-hosted Go reverse proxy with HTTP/HTTPS and TCP/UDP forwarding, auto ACME certs, multi-node agents, Relay tunnels, and traffic quotas — managed via web UI.
oc8oc8 is a self-hosted, open-source platform for orchestrating AI agents that work across business software using the Model Context Protocol with human oversight.
octo-serverSelf-hosted Go backend for OCTO — humans and AI agents collaborate via REST/WebSocket, Lobster agents, and WuKongIM messaging.
OmniKey AI: Unified Key ManagerOmniKey AI is a self-hosted, open-source LLM proxy routing requests across 12 free providers through unified OpenAI and Gemini endpoints with automatic failover.
openaseOpenASE is a self-hosted, open-source platform turning tickets into working code via AI agents with full traceability and a Kanban control plane in a single Go binary.
OpengeniOpengeni is an open-source, self-hostable agentic service for AI agents with durable sessions and human approvals.
openrunA self-hosted GitOps platform that deploys web apps to Docker or Kubernetes using declarative config, with built-in SSO, RBAC, and database bindings.
OpenshipSelf-hosted open-source PaaS: hand it a repo, a folder or an artifact and it detects your stack, builds it, then serves it on your domain with free certificates.
osInfinite OS is a self-hosted, open-source hosting environment that runs in a single container with a cPanel-like dashboard to deploy apps without Dockerfiles.
OxMgrSelf-hosted Rust process manager replacing PM2 — runs any executable on Linux, macOS, Windows with health checks, resource limits, and a live terminal dashboard.
PhreshOSA self-hosted platform that runs web apps on your hardware and serves a browser desktop for people and AI agents to use them together.
pm2-hawkeyeSelf-hosted, open-source dashboard to monitor PM2 processes with live CPU/memory metrics, log streaming, one-click restart, and alerting via webhooks or ntfy.
0 comments