Open-source task management for the agentic era.
The command center for solo entrepreneurs who delegate work to AI agents.

AI coding agents (Claude Code, Cursor, Windsurf) are powerful executors — but managing multiple agents across multiple projects is chaos. There's no shared task board, no inbox, no way to see who's working on what or whether they finished.

Mission Control gives your AI agents structure. Agents get roles, inboxes, and reporting protocols. You delegate work through a visual dashboard, they execute and report back. You stay in control without micromanaging.

How is this different from Linear, Asana, or Notion? Mission Control was built agent-first. Agents read and write tasks through a token-optimized API, report progress to your inbox, and ask you for decisions. You manage outcomes, not keystrokes. And it runs locally — no cloud dependency, no API keys, no vendor lock-in.

• Eisenhower Matrix — Prioritize by importance and urgency with drag-and-drop between quadrants
• Kanban Board — Track work through Not Started, In Progress, and Done columns
• Goal Hierarchy — Long-term goals with milestone tracking, progress bars, and linked tasks
• Brain Dump — Capture ideas instantly, triage into tasks later
• Agent Crew — 5 built-in agents + create unlimited custom agents with unique instructions
• Skills Library — Define reusable knowledge modules and inject them into agent prompts
• Multi-Agent Tasks — Assign a lead agent + collaborators for team-based work
• Orchestrator — Run /orchestrate to spawn all agents on pending work simultaneously
• Autonomous Daemon — Background process that automatically polls tasks, spawns Claude Code sessions, enforces concurrency, and provides a real-time dashboard
• One-Click Execution — Press play on any task card to spawn a Claude Code session; live status indicators, success/failure toasts, and automatic completion (task → done, inbox report, activity log)
• Token-Optimized API — Filtered queries, sparse field selection, 92% context compression (~50 tokens vs ~5,400)
• Inbox & Decisions — Full agent communication layer: delegation, reports, questions, and approvals
• Cmd+K Search — Global search across tasks, projects, goals, and brain dump entries
• Error Resilience — Error boundaries on every page with retry buttons, plus global error handler for crash recovery
• API Pagination — All 9 GET endpoints support limit and offset with a meta object (total, filtered, returned)
• 193 Automated Tests — Vitest suite covering validation schemas, data layer operations, and full agent communication flow
• Skills Injection — Skills from the library are embedded into agent command files bidirectionally (agent→skill and skill→agent)
• Accessibility — ARIA live regions for drag-and-drop screen reader announcements, focus trapping on detail panels
• CI Pipeline — GitHub Actions runs typecheck, lint, build, and tests on every push and PR

The web UI works standalone for task management, prioritization, and goal tracking. Claude Code is needed to execute tasks via agents. Any AI coding tool that can access local files (Cursor, Windsurf, etc.) can also participate — see Works With below.

git clone https://github.com/MeisnerDan/mission-control.git
cd mission-control/mission-control   # repo folder → app folder (where package.json lives)
pnpm install
pnpm dev

Open http://localhost:3000 and click "Load Demo Data" to see it in action with sample tasks, agents, and messages.

1. Explore the dashboard — see task counts, agent workloads, and recent activity at a glance
2. Drag tasks on the Priority Matrix — move tasks between Do, Schedule, Delegate, and Eliminate
3. Click a task card to open the detail panel — edit description, subtasks, and acceptance criteria
4. Click the 🚀 Launch button on a task assigned to an agent — spawns a Claude Code session that executes the work (requires Claude Code)
5. Open Claude Code in this workspace and run /daily-plan to see slash commands in action

Mission Control stores all data in local JSON files. No database, no cloud dependency. AI agents interact by reading and writing these files — the same source of truth the web UI uses.

1. You create a task          ──>  Assign to an agent role (e.g., Researcher)
2. Press play (or daemon)     ──>  Spawns a Claude Code session with agent persona
3. Agent executes             ──>  Does the work, updates progress
4. Mission Control completes  ──>  Auto-marks done, posts report, logs activity
5. You review                 ──>  Read reports in inbox, answer questions

Multiple agents can work in parallel across different tasks. The orchestrator (/orchestrate) automates this loop for all pending tasks at once. The daemon (pnpm daemon:start) takes this further — it runs 24/7 as a background process, automatically polling for new tasks and dispatching them to agents on a configurable schedule.

Mission Control includes 193 automated tests across 3 suites:

pnpm test        # Run all tests
pnpm check       # Typecheck + lint
pnpm verify      # Full verification: typecheck + lint + build + test

Every API endpoint is designed for minimal token consumption. Your agents spend tokens doing work, not parsing bloated payloads.

# Get only your in-progress tasks (~50 tokens vs ~5,400 for everything)
GET /api/tasks?assignedTo=developer&kanban=in-progress

# Sparse fields — return only what you need
GET /api/tasks?fields=id,title,kanban

# Get just the DO quadrant (important + urgent)
GET /api/tasks?quadrant=do

# Paginated results with metadata
GET /api/tasks?limit=10&offset=0
# → { data: [...], meta: { total: 47, filtered: 47, returned: 10, limit: 10, offset: 0 } }

# Compressed context — entire workspace state in ~650 tokens
# (vs ~10,000+ for raw JSON files)
pnpm gen:context  # outputs data/ai-context.md

# Run a single task — spawns a Claude Code session
POST /api/tasks/:id/run

# Run all eligible tasks in a project (respects concurrency limits)
POST /api/projects/:id/run

# Get live status of all active runs
GET /api/runs

All write endpoints use Zod validation (malformed data returns field-level errors) and async-mutex locking (concurrent writes from multiple agents queue safely, never corrupt data).

Agents are fully editable — change their name, instructions, capabilities, and linked skills at any time through the Crew UI or by editing data/agents.json directly.

Run these in any Claude Code session opened in this workspace:

pnpm daemon:start    # Start the autonomous daemon (background process)
pnpm daemon:stop     # Stop the daemon gracefully
pnpm daemon:status   # Show daemon status, active sessions, and stats

The daemon runs as a background Node.js process, polling tasks.json for pending work and spawning Claude Code sessions via claude -p. It enforces concurrency limits, retries failed tasks, and runs scheduled commands (daily-plan, standup, weekly-review) on cron schedules. Monitor everything from the /daemon dashboard.

Note on authentication: The daemon spawns Claude Code directly via claude -p — it does not extract or transmit OAuth tokens, make raw API calls, or use the Agent SDK. Your Claude account credentials stay within Claude Code's own authentication layer, the same as running claude -p from your terminal. This is local automation of an official Anthropic product, not a third-party integration.

mission-control/              Next.js 15 web app (the visual interface)
mission-control/data/          JSON data files (the shared source of truth)
  tasks.json                   Tasks with Eisenhower + Kanban + agent assignment
  goals.json                   Long-term goals and milestones
  projects.json                Projects with team members
  agents.json                  Agent registry (profiles, instructions, capabilities)
  skills-library.json          Reusable knowledge modules for agents
  inbox.json                   Agent <-> human messages and reports
  decisions.json               Pending decisions requiring human judgment
  activity-log.json            Timestamped event log of all activity
  ai-context.md                Generated ~650-token workspace snapshot
  daemon-config.json           Daemon configuration (schedule, concurrency, etc.)
  daemon-status.json           Daemon runtime state (sessions, history, stats)
  active-runs.json             Live task execution tracking (status, PIDs, errors)
mission-control/scripts/daemon/ Autonomous agent daemon (node-cron + claude -p)
mission-control/__tests__/     Automated tests (validation, data, integration, daemon)
.claude/commands/              Auto-generated slash commands per agent
scripts/                       Orchestration scripts (tmux parallel agents)
docs/                          Business plans and strategies

• Local-first — No database, no cloud, no API keys. Your data stays on your machine in plain JSON files.
• JSON as IPC — Humans (web UI) and agents (file reads + API) share the same source of truth. No sync layer needed.
• BYOAI — Works with any agent that can read files: Claude Code, Cursor, Windsurf, or a custom script.
• Zod + Mutex — All API writes are validated with Zod schemas and serialized with async-mutex to prevent data corruption during concurrent multi-agent operations.

Mission Control runs locally and integrates with AI coding tools through the filesystem and CLI:

• Claude Code — Open this workspace in Claude Code to use slash commands (/orchestrate, /daily-plan, /standup, etc.) and let agents read/write task data directly.
• Claude Cowork — Cowork agents can use Mission Control as a tool by reading the workspace's CLAUDE.md and JSON data files directly — no special plugin required.
• Any file-aware agent — Cursor, Windsurf, or custom scripts can read the JSON data files and call the API endpoints to participate in the agent loop.