Mission Control is a self-hosted dashboard for operating fleets of AI agents. It centralizes what would otherwise be scattered tooling — a Kanban task board with quality-review gates, agent lifecycle management, real-time activity feeds, token/cost tracking, a memory knowledge graph, and security auditing — behind a single Next.js app backed only by SQLite, so it runs with no Redis, Postgres, or other external services. Updates stream over WebSocket and SSE, and access is role-based (viewer/operator/admin) with session, API-key, and Google Sign-In auth.
It is framework-agnostic via an adapter layer that normalizes registration, heartbeats, and task reporting for OpenClaw, CrewAI, LangGraph, AutoGen, and the Claude SDK, and can connect to multiple agent gateways at once — or run gateway-free for task and project management. Notable capabilities include:
- a Skills Hub that browses and installs agent skills from registries (ClawdHub, skills.sh), scanning them for prompt injection, credential leaks, and dangerous shell commands before install
- a four-layer agent eval framework (output scoring, trace/loop detection, tool-reliability metrics, drift detection) plus per-agent trust scoring and secret detection
- natural-language recurring tasks ("every morning at 9am") converted to cron with template-spawned child tasks
- a read-only Claude Code bridge surfacing local sessions, team tasks, and cost estimates
- outbound webhooks with retries and HMAC signing, GitHub Issues sync, and inter-agent messaging
The project labels itself alpha: APIs, schemas, and config formats may change between releases, and the docs stress hardening (TLS, host allowlists, changed credentials) before any network-exposed deployment. The REST surface is documented via an OpenAPI 3.1 spec with roughly a hundred endpoints.
