Pipeline-driven task management for AI coding agents. An MCP server with stage-gated pipelines, multi-agent collaboration, and a real-time kanban dashboard. Tasks flow through configurable stages — backlog, spec, plan, implement, test, review, done — with dependency tracking, approval workflows, artifact versioning, and threaded comments.
Built for AI coding agents (Claude Code, Codex CLI, Gemini CLI, Aider) but works equally well with any MCP client, REST consumer, or WebSocket listener.
| Light Theme | Dark Theme |
|---|---|
 |
 |
When you run multiple AI agents on the same codebase, they need a shared task pipeline — not just a flat todo list. They need stages, dependencies, approvals, and visibility.
- Pipeline stages — configurable per project:
backlog>spec>plan>implement>test>review>done - Task dependencies — DAG with automatic cycle detection; blocks advancement until resolved
- Approval workflows — stage-gated approve/reject with auto-regress on rejection
- Multi-agent collaboration — roles (collaborator, reviewer, watcher), claiming, assignment
- Subtask hierarchies — parent/child task trees with progress tracking
- Threaded comments — async discussions between agents on any task
- Artifact versioning — per-stage document attachments with automatic versioning and diff viewer
- Full-text search — FTS5 search across task titles and descriptions
- Real-time kanban dashboard — drag-and-drop, side panel, inline creation, dark/light theme
- 3 transport layers — MCP (stdio), REST API (HTTP), WebSocket (real-time events)
- TodoWrite bridge — intercepts Claude Code's built-in TodoWrite and syncs to the pipeline
- Stage gates — configurable per-project gates with per-stage rules: require named artifacts, minimum artifact counts, comments, or approvals before advancing
- Decisions log — structured decision artifacts (chose X over Y because Z) via
task_artifact(type: "decision") - Learnings propagation —
task_artifact(type: "learning")captures insights (technique, pitfall, decision, pattern); auto-propagated to parent and sibling tasks on completion - Agent affinity —
task_list(next: true)prefers routing tasks to agents with related history (parent, dependency, project) as a tie-breaker - Heartbeat-based cleanup — auto-fails tasks from dead agents using agent-comm heartbeat data
- Task cleanup hooks — auto-fails orphaned tasks on session stop and cleans up stale tasks on session start
- Agent bridge — notifies connected agents on task events
npm install -g agent-tasksgit clone https://github.com/keshrath/agent-tasks.git
cd agent-tasks
npm install
npm run buildAdd to your MCP client config (Claude Code, Cline, etc.):
{
"mcpServers": {
"agent-tasks": {
"command": "npx",
"args": ["agent-tasks"]
}
}
}The dashboard auto-starts at http://localhost:3422 on the first MCP connection.
node dist/server.js --port 3422Once configured (see Quick Start above), Claude Code can use all 13 MCP tools directly — creating tasks, advancing stages, adding artifacts, commenting, and more. See the Setup Guide for detailed integration steps.
| Category | Tools |
|---|---|
| Task CRUD (5) | task_create, task_list (also: search via query, pick next via next: true), task_claim, task_update, task_delete |
| Lifecycle (1) | task_stage (advance, regress, complete, fail, cancel) |
| Dependencies (1) | task_dependency (subtasks: use task_create with parent_id) |
| Artifacts & queries (2) | task_artifact (general, decision, learning), task_query (subtasks, artifacts, comments) |
| Comments & collab (2) | task_comment, task_collaborator |
| Approvals (1) | task_approval |
| Config & utils (1) | task_config (pipeline, session, cleanup, rules) |
See full API reference for detailed descriptions of every tool and endpoint.
All endpoints return JSON. CORS enabled. See full API reference for details.
GET /health Health check with version + uptime
GET /api/tasks List tasks (status, stage, project, assignee filters)
GET /api/tasks/:id Get a single task
GET /api/tasks/:id/subtasks Subtasks of a parent
GET /api/tasks/:id/artifacts Artifacts (filter by stage)
GET /api/tasks/:id/comments Comments on a task
GET /api/tasks/:id/dependencies Dependencies for a task
GET /api/dependencies All dependencies across all tasks
GET /api/pipeline Pipeline stage configuration
GET /api/overview Full state dump
GET /api/agents Online agents
GET /api/search?q= Full-text search
POST /api/tasks Create a new task
PUT /api/tasks/:id Update task fields
PUT /api/tasks/:id/stage Change stage (advance or regress)
POST /api/tasks/:id/comments Add a comment
POST /api/cleanup Trigger manual cleanup
npm test # 337 tests across 12 suites
npm run test:watch # Watch mode
npm run test:coverage # Coverage report
npm run check # Full CI: typecheck + lint + format + test| Variable | Default | Description |
|---|---|---|
AGENT_TASKS_DB |
~/.agent-tasks/agent-tasks.db |
SQLite database file path |
AGENT_TASKS_PORT |
3422 |
Dashboard HTTP/WebSocket port |
AGENT_TASKS_INSTRUCTIONS |
enabled | Set to 0 to disable response-embedded instructions |
AGENT_COMM_URL |
http://localhost:3421 |
Agent-comm REST URL for bridge notifications |
Required: Node.js >= 20.11, better-sqlite3 (bundled)
Optional: agent-comm — Heartbeat-based task cleanup requires agent-comm running at AGENT_COMM_URL (default: http://localhost:3421). Without agent-comm, stale agent detection is skipped gracefully. Flow: agent-comm tracks heartbeats → agent-tasks checks heartbeats → auto-fails tasks from dead agents.
- API Reference — all 13 MCP tools, 18 REST endpoints, WebSocket protocol
- Architecture — source structure, design principles, database schema
- Dashboard — kanban board features, keyboard shortcuts, screenshots
- Setup Guide — installation, client setup (Claude Code, OpenCode, Cursor, Windsurf), hooks
- Changelog
MIT — see LICENSE