Mission Control

Mission Control is a self-hosted control plane for operating AI agents. It sits above agent runtimes rather than replacing their reasoning or tool loops, giving one local dashboard for dispatching tasks, inspecting runs, reviewing failures, tracking spend, and coordinating runtimes, with state in SQLite. The project labels itself alpha software whose APIs, schemas, and configuration may change between releases.

The surface is organised by area: tasks move through inbox, assignment, execution, review, an Aegis quality gate, and completion, and an approval record is required before a task reaches done. Agent views combine registration, presence, sessions, runtime adapters, configuration, and workspace sync. Operations covers an activity stream, schedules, alerts, webhooks, logs, token use, and cost views; knowledge covers a memory browser, relationship graph, and skills registry with local synchronization; governance covers roles, API keys, security events, approvals, audits, and evals.

Adapters and observation surfaces cover OpenClaw, Claude Code, Codex, CrewAI, LangGraph, AutoGen, and Claude SDK workflows, with adapter depth varying by runtime. Agents connect over the REST API with a bearer API key — register an agent, create a task, and let the agent claim its queue — or through the CLI, an MCP server, WebSocket, and server-sent events. The REST contract is described by an OpenAPI document that a running instance serves alongside an interactive reference.

It runs from source with Node.js 22 and pnpm, or from a published multi-architecture container image with a hardened Compose overlay for network-accessible deployments. Runtime data defaults to a local directory that can be moved with an environment variable. The security guidance is explicit: keep the instance on a trusted network unless a TLS reverse proxy and an allowed-hosts list are configured, replace generated credentials before wider access, and treat agent messages, skill packages, webhooks, and MCP content as untrusted input.

The project also states when it fits poorly — a single agent on one machine that stays legible from its own CLI, a need for managed multi-tenant hosting, a wish for the tool to define planning and tool use, or a deployment that cannot absorb alpha schema and API changes.

Features

• Task pipeline: inbox, assignment, execution, review, quality gate, and completion receipts, with approval required before done
• Agent registry: registration, heartbeats, sessions, runtime discovery, configuration, and workspace files in one view
• Runtime adapters: OpenClaw, Claude Code, Codex, CrewAI, LangGraph, AutoGen, and Claude SDK workflows
• Interfaces: web UI, CLI, MCP server, OpenAPI-described REST API, WebSocket, and server-sent events
• Cost and token tracking: per-agent spend views alongside alerts and webhooks
• Memory and skills: a memory browser with a relationship graph, plus a skills hub that scans registry content before installation
• Schedules: recurring task templates that create dated work on a cron schedule
• Governance: roles, API keys, security events, trust signals, approvals, audits, and evals
• Deployment paths: source install, Docker Compose, a published container image, and a hardened overlay for networked use
• Local state: SQLite through better-sqlite3 in write-ahead-log mode, with a configurable data directory