Pipeline task management for AI agents - stages, dependencies, artifacts, claiming
io.github.keshrath/agent-tasks โ MCP Server
The io.github.keshrath/agent-tasks Model Context Protocol (MCP) server provides pipeline task management for AI agents. It supports stage-gated pipelines with dependencies and artifacts, enabling multi-agent collaboration and a real-time kanban dashboard. The project is implemented in TypeScript and uses SQLite, with WebSocket support.
๐ ๏ธ Key Features
Pipeline-driven task management
Stage-gated pipelines
Dependencies and artifacts
Multi-agent collaboration
Real-time kanban dashboard
๐ Use Cases
Managing AI coding agent workflows through pipeline stages
Coordinating multi-agent collaboration on shared tasks
Tracking task progress in real time via kanban
โก Developer Benefits
Provides MCP tooling (listed as MCP tools)
Exposes REST endpoints (listed as REST endpoints)
Built with TypeScript, using SQLite and WebSocket
โ ๏ธ Limitations
Description does not specify authentication, deployment options, or the exact tool/endpoint APIs.
Pipeline-driven task management for AI coding agents. An MCP server with stage-gated pipelines, multi-agent collaboration, and a native Claude Code /pipeline pane. Tasks flow through configurable stages โ backlog, spec, plan, implement, test, review, done โ with dependency tracking, stage gates, 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 or REST consumer.
/pipeline opens the board beside a Claude Code conversation; Tab and Enter open a task
The user manual shows every view and action of the pane, with screenshots and short clips from the real CLI.
2.0 replaced the web dashboard with the /pipeline pane inside Claude Code; there is no browser UI, WebSocket or agent-desk plugin any more.
From 1.x
The dashboard on http://localhost:3422 is gone: open the board with /pipeline in Claude Code (install the plugin, below). The agent-desk Tasks tab is gone with it.
REST: GET /api/overview and PUT /api/tasks/:id/stage are gone (use GET /api/board and POST /api/tasks/:id/stage with {"action": "advance"}), and so are the routes nothing called; see the REST API.
A 1.x process still holding port 3422 is named by the pane (An older agent-tasks (1.10.13) holds the port) and in the status line (old server 1.10.13 on the port): end the Claude Code sessions still running it (or kill the old agent-tasks process) so a 2.x process takes the port, then Refresh.
Why agent-tasks?
When you run multiple AI agents on the same codebase, they need a shared task pipeline โ not just a flat todo list. They need stages, dependencies, stage gates, and visibility.
Features
Pipeline stages โ configurable per project: backlog > spec > plan > implement > test > review > done
Task dependencies โ DAG with automatic cycle detection; blocks advancement until resolved
Multi-agent claiming โ claim, assign, send back to an earlier stage with a reason
Subtask hierarchies โ parent/child task trees with progress tracking
Threaded comments โ async discussions between agents on any task
Artifact versioning โ per-stage document attachments with automatic versioning
Full-text search โ FTS5 search across task titles and descriptions
Claude Code /pipeline pane โ the board beside the conversation, keyboard-first: tasks by stage, task detail with artifacts and comments, advance, send back, complete, fail, cancel, comment and delete, a status line entry
2 transport layers โ MCP (stdio) and a loopback REST API (HTTP) that the pane reads
TodoWrite bridge โ intercepts Claude Code's built-in TodoWrite and syncs to the pipeline
Stage gates โ configurable per-project gates with per-stage rules: require named artifacts, minimum artifact counts or a comment before advancing (the schema also has require_approval, but no MCP tool or REST route records an approval, so a stage with it set cannot be left; keep it off)
Decisions log โ structured decision artifacts (chose X over Y because Z) via task_artifact(type: "decision")
Learnings propagation โ task_artifact(type: "learning") captures insights (technique, pitfall, decision, pattern); auto-propagated to parent and sibling tasks on completion
Agent affinity โ task_list(next: true) prefers routing tasks to agents with related history (parent, dependency, project) as a tie-breaker
Heartbeat-based cleanup โ auto-fails tasks from dead agents using agent-comm heartbeat data
Task cleanup hooks โ auto-fails orphaned tasks on session stop and cleans up stale tasks on session start
The first MCP process to start also serves the REST API on http://localhost:3422; the others retry every 5 s and take over when it exits.
Option 2: Claude Code plugin (MCP server + /pipeline pane)
bash
claude plugin marketplace add keshrath/agent-tasks
claude plugin install agent-tasks@agent-tasks
The plugin runs the MCP server and adds a Claude Code mod (function hooks, Claude Code 2.1.289+): /pipeline opens the board beside the conversation. The tutorial walks through it. Remove any hand-written agent-tasks entry from ~/.claude.json so tools do not appear twice.
Option 3: Standalone REST server
bash
node dist/server.js --port 3422
Claude Code Integration
Once configured (see Quick Start above), Claude Code can use all 8 MCP tools directly โ creating tasks, advancing stages, adding artifacts, commenting, and more. With the plugin installed, /pipeline opens the board in a pane beside the conversation (Claude Code 2.1.289+). See the Setup Guide for detailed integration steps.
See full API reference for detailed descriptions of every tool and endpoint.
REST API (8 endpoints)
All endpoints return JSON. Loopback only: a foreign Host gets 403, so does any request carrying an Origin header (no browser calls this API), and request bodies must be application/json. Everything else goes through the MCP tools. See full API reference for details.
code
GET /health Health check with version + uptime
GET /api/tasks/:id One task, as task_get (?include=subtasks,comments,transitive_deps)
GET /api/pipeline Pipeline stage configuration
GET /api/board Open and recently closed tasks (what the pane polls)
POST /api/tasks Create a task
POST /api/tasks/:id/stage Lifecycle action (claim, assign, advance, regress, complete, fail, cancel)
POST /api/tasks/:id/comments Add a comment
DELETE /api/tasks/:id Delete a task
Testing
bash
npm test# vitest: 472 tests across 25 files
npm run test:watch # Watch mode
npm run test:coverage # Coverage report
npm run check # Full CI: typecheck + lint + format + test
npm run plugin:check # the /pipeline pane: claude plugin validate + its tests (needs the claude CLI)
npm run e2e:claude # the pane in the real Claude Code CLI (see CONTRIBUTING.md)
Environment variables
Variable
Default
Description
AGENT_TASKS_DB
~/.agent-tasks/agent-tasks.db
SQLite database file path
AGENT_TASKS_PORT
3422
REST API port (the /pipeline pane reads it)
AGENT_TASKS_HOST
127.0.0.1
REST bind address (0.0.0.0 exposes it to the network)
AGENT_TASKS_USER
the OS user
Who the /pipeline pane acts as (created_by, comment author, by)
AGENT_TASKS_INSTRUCTIONS
enabled
Set to 0 to disable response-embedded instructions
AGENT_TASKS_RETENTION_DAYS
30
Days the cleanup keeps resolved approval records
AGENT_TASKS_COMPLETED_RETENTION_DAYS
7
Days a completed task is kept (and listed under the pane's Done) before purge
AGENT_TASKS_FAILED_RETENTION_DAYS
7
Days a failed task is kept before purge
AGENT_TASKS_CANCELLED_RETENTION_DAYS
1
Days a cancelled task is kept before purge
AGENT_TASKS_ORPHAN_TIMEOUT_MINUTES
360
Minutes before an in-progress task with no assignee is cancelled
AGENT_TASKS_COMM_TIMEOUT_MS
5000
Timeout of the cleanup's agent-comm requests
AGENT_TASKS_DEAD_SESSION_GRACE_MINUTES
60
task-cleanup-start.js: idle minutes before a dead session's task is failed
AGENT_TASKS_HEARTBEAT_FRESH_MINUTES
15
task-cleanup-start.js: an agent with a heartbeat this recent counts as alive
Optional (soft dependencies โ fail-open, HTTP-only, no npm dep):
agent-comm โ Heartbeat-based task cleanup and event notifications. agent-comm tracks heartbeats โ agent-tasks checks heartbeats โ auto-fails tasks from dead agents. Also sends direct messages on claim/advance and posts to a channel on comments. Without agent-comm, stale agent detection and notifications are skipped gracefully.
agent-knowledge โ Knowledge persistence for task learnings and decisions. On task completion, the KnowledgeBridge pushes learning and decision artifacts to agent-knowledge via POST /api/knowledge. Entries are auto-indexed with embeddings, auto-linked to similar entries, and git-synced. Without agent-knowledge, artifacts stay in agent-tasks only.
Documentation
Tutorial โ the /pipeline pane step by step, recorded from the real CLI
User Manual โ the /pipeline pane view by view (screenshots and GIFs), every tool, the pipeline, gates, cleanup
API Reference โ all 8 MCP tools and the REST endpoints
Hooks โ the Claude Code command hooks scripts/setup.js installs