AgentPlanner MCP Server

MCP server for AgentPlanner — AI agent orchestration with planning, dependencies, knowledge graphs, and human oversight. Works with Claude Desktop, Claude Code, ChatGPT, Cursor, Windsurf, Cline, and any MCP-compatible client.

Prerequisites

• An AgentPlanner account at agentplanner.io
• An API token (Settings > API Tokens in the AgentPlanner UI)

Setup

Claude Desktop — one-click install (.mcpb)

The fastest path. Download agent-planner.mcpb from the latest release, double-click it, and Claude Desktop will install the extension and prompt for your AgentPlanner API token. No Node.js setup, no JSON editing.

To build the bundle yourself:

npm run build:mcpb        # produces agent-planner.mcpb
npm run validate:mcpb     # schema-check manifest.json

Manual config (Claude Desktop, Claude Code, Cursor, etc.)

Add to your MCP client config (claude_desktop_config.json, .cursor/mcp.json, etc.):

{
  "mcpServers": {
    "agentplanner": {
      "command": "npx",
      "args": ["-y", "agent-planner-mcp"],
      "env": {
        "API_URL": "https://agentplanner.io/api",
        "USER_API_TOKEN": "your_token_here"
      }
    }
  }
}

Thin local client (v1)

A lightweight CLI loop for task-driven workflows. No MCP client required — useful when an agent (Claude Code, OpenClaw, a script) just needs to read its current task as files and write status back.

Mental model

• AgentPlanner (the API) is the source of truth.
• .agentplanner/ files are a regeneratable cache, written by the CLI for the agent to read.
• The agent works in the real repo. Status changes flow back via explicit writeback commands. There is no live sync.

Running locally? See agent-planner/LOCAL_QUICKSTART.md for the 5-minute path to a full local stack you can point this CLI at. Use --api-url http://localhost:3000 in the login step below.

The loop

# 1. Login — saves credentials and auto-selects a default plan
#    (pass --plan-id to pick one, or it auto-selects if you have exactly one plan)
npx agent-planner-mcp login --token <token> --api-url https://agentplanner.io/api [--plan-id <id>]
#    Localhost variant (after `docker compose -f docker-compose.local.yml up`):
npx agent-planner-mcp login --token <token> --api-url http://localhost:3000

# 2. See your task queue
npx agent-planner-mcp tasks [--plan-id <id>]

# 3. Pick the next task and pull context (claims it for 30 minutes)
npx agent-planner-mcp next [--plan-id <id>]
#    Force a fresh recommendation even if you have active work:
npx agent-planner-mcp next --fresh

# 4. Or pull context for a specific plan/node (no claim, no status change)
npx agent-planner-mcp context --plan-id <plan-id> --node-id <node-id>
#    If a default plan is set, --plan-id can be omitted:
npx agent-planner-mcp context --node-id <node-id>

# 5. Explicit writeback. No live sync.
npx agent-planner-mcp start                          # claim + mark in_progress
npx agent-planner-mcp blocked --message "Waiting on API decision"
npx agent-planner-mcp done    --message "Implemented and verified"

next resolution order

next is a smart picker. It resolves in this order:

1. Resume — if any task in scope is in_progress, pick it. (Source: resume_in_progress.)
2. Recommend — call suggest_next_tasks (dependency- and RPI-aware) for a fresh pick. (Source: suggest_next_tasks.)
3. Fallback — first not_started task in your queue. (Source: my_tasks_fallback.)

tasks is the queue view; next is the smart picker; next --fresh skips step 1 and forces a fresh recommendation even when active work exists.

What start, blocked, done actually do

| Command | Status | Claim | Log entry | Learning written to Graphiti | |---|---|---|---|---| | start | in_progress | claim (30m TTL) | — | — | | blocked --message ... | blocked | release | challenge | — | | done --message ... | completed | release | progress | yes (entry_type: learning) |

All hooks are best-effort: claim/release/learning failures do not block the status update. Claim collisions (another agent already holds the lease) are reported but not fatal.

What current-task.md surfaces

Beyond title, description, agent_instructions, and acceptance criteria, the generated current-task.md includes BDI signals from the API responses already being fetched:

• Plan health — quality_score, rationale, coherence_checked_at (or "never")
• Coherence warning — flagged when node.coherence_status is contradiction_detected or stale_beliefs, with concrete next-step pointers (check_contradictions, recall_knowledge)
• Detected contradictions — listed when present in the node context
• Task mode — shown when not free (RPI awareness for research/plan/implement)
• Linked goals, relevant knowledge (top 5), plan progress snapshot

When to use CLI vs MCP vs API skill

| You want… | Use | |---|---| | Zero-setup local task context for any coding agent (Claude Code, OpenClaw, scripts) | CLI (this thin client) | | Rich, structured tool access from inside an MCP-aware agent (Claude Desktop, Cursor, etc.) | MCP (run npx agent-planner-mcp as an MCP server) | | Direct programmatic integration from your own service | API (REST endpoints; same routes the MCP and CLI use) |

The CLI is intentionally thin: it covers the read context + writeback loop and nothing else. For decomposition, dependency creation, knowledge graph queries, RPI chains, coherence runs, and goal management, use the MCP server (or the API directly).

Agent Loop Facade

AgentPlanner API now exposes a narrow /agent/* facade for the main autonomous loop. MCP uses this facade when available and falls back to older domain endpoints for self-hosted older APIs.

Primary mappings:

| MCP tool | Preferred API endpoint | |---|---| | briefing | GET /agent/briefing | | claim_next_task | POST /agent/work-sessions | | update_task with session_id + completed | POST /agent/work-sessions/:id/complete | | update_task with session_id + blocked | POST /agent/work-sessions/:id/block | | form_intention | POST /agent/intentions when available, with domain-endpoint fallback |

Validation:

npm run validate:mcp-loop

This checks that the MCP tools route through the facade for briefing, task claim/start, and session completion/blocking.

Claude Desktop

Add to your claude_desktop_config.json:

{
  "mcpServers": {
    "agent-planner": {
      "command": "npx",
      "args": ["-y", "agent-planner-mcp"],
      "env": {
        "USER_API_TOKEN": "your-token",
        "API_URL": "https://agentplanner.io/api"
      }
    }
  }
}

Config location: ~/Library/Application Support/Claude/claude_desktop_config.json (macOS) | %APPDATA%\Claude\claude_desktop_config.json (Windows)

Claude Code

claude mcp add agent-planner -- npx -y agent-planner-mcp

Then set the env vars USER_API_TOKEN and API_URL=https://agentplanner.io/api.

ChatGPT

1. Settings > Apps > Advanced > Developer mode
2. Add MCP Server > URL: https://agentplanner.io/mcp
3. Auth type: API Key > enter your token from agentplanner.io Settings

Cursor

Add to .cursor/mcp.json in your project root:

{
  "mcpServers": {
    "agent-planner": {
      "command": "npx",
      "args": ["-y", "agent-planner-mcp"],
      "env": {
        "USER_API_TOKEN": "your-token",
        "API_URL": "https://agentplanner.io/api"
      }
    }
  }
}

Windsurf

Add to ~/.codeium/windsurf/mcp_config.json:

{
  "mcpServers": {
    "agent-planner": {
      "command": "npx",
      "args": ["-y", "agent-planner-mcp"],
      "env": {
        "USER_API_TOKEN": "your-token",
        "API_URL": "https://agentplanner.io/api"
      }
    }
  }
}

Cline (VS Code)

Add the same JSON config to your Cline MCP settings in VS Code.

Any HTTP MCP Client

• Endpoint: https://agentplanner.io/mcp
• Discovery: https://agentplanner.io/.well-known/mcp.json
• Auth header: Authorization: ApiKey <your-token>
• Transport: Streamable HTTP (MCP 2025-03-26)

Key Features

• 24 BDI-aligned tools for state, goals, and committed actions — no CRUD shapes, every tool answers a whole agentic question
• Full mutation surface (v1.0) — agents and humans-via-agents can manage every plan/node/org property without leaving the conversation; UI is optional inspection
• Draft-status seam — autonomous agent creation lands as drafts surfacing in the dashboard pending queue; human-directed creation defaults to active
• Dependency graph — cycle detection, impact analysis, critical path
• Progressive context — 4-layer context assembly with token budgeting
• Knowledge graph — temporal knowledge via Graphiti (entities, facts, contradictions)
• RPI chains — Research → Plan → Implement task decomposition (one-call shortcut)
• Task claims — TTL-based locking for multi-agent coordination
• Organizations — multi-tenant isolation with member management

Available Tools (v1.0.0)

Beliefs (read state)

• briefing — bundled mission control state in one call
• task_context — single task at progressive depth 1-4
• goal_state — single goal deep-dive (details + quality + progress + bottlenecks + gaps)
• recall_knowledge — knowledge graph query (facts, entities, episodes, contradictions)
• search — text search across plans/nodes
• plan_analysis — impact, critical path, bottlenecks, coherence

Desires (goals)

• list_goals — goals with health rollup
• update_goal — atomic goal update (subsumes link/unlink/achievers)
• derive_subgoal (v1.0) — propose a sub-goal under an existing parent

Intentions — execution

• claim_next_task — pick + claim + load context (one call)
• update_task — atomic status + log + claim release + learning
• release_task — explicit handoff
• queue_decision — escalate to human (real decision queue)
• resolve_decision — pick up human's answer (atomically materializes any proposed_subtasks)
• add_learning — record knowledge episode

Intentions — creation (v1.0)

• form_intention — create plan + initial tree under a goal, atomically
• extend_intention — add children under an existing parent (lightweight)
• propose_research_chain — RPI triple with 2 blocking edges, in one call

Intentions — structural mutation (v1.0)

• update_plan — edit any plan property
• update_node — edit any node property except status
• move_node — reparent within plan; cycle-safe
• link_intentions / unlink_intentions — manage dependency edges
• delete_plan / delete_node — soft-delete via status='archived' (recoverable)

Intentions — sharing & collaboration (v1.0)

• share_plan — atomic visibility + add/remove collaborators
• invite_member — add user to org (by user_id or email)
• update_member_role — owner-only role change
• remove_member — owner/admin removes non-owner member

Utility

• get_started — dynamic reference for new agents

See SKILL.md for full descriptions, the human-steering scenarios (A/B/C), and status='draft' vs status='active' guidance.

LLM Skill Reference

See SKILL.md for a complete reference designed to be consumed by LLMs. Include it in system prompts or agent configurations to give any LLM full knowledge of how to use AgentPlanner tools effectively.

See AGENT_GUIDE.md for a quick reference card.

Transport Modes

stdio (default)

For local use with Claude Desktop, Claude Code, Cursor, Windsurf, Cline:

npx agent-planner-mcp

HTTP/SSE

For remote access (ChatGPT, cloud deployments, multi-agent systems):

MCP_TRANSPORT=http npx agent-planner-mcp
# Listens on http://127.0.0.1:3100

Production endpoint: https://agentplanner.io/mcp

See HTTP_MODE.md for details.

Local Development

git clone https://github.com/TAgents/agent-planner-mcp.git
cd agent-planner-mcp
npm install
npm run setup    # Interactive setup wizard
npm run dev      # Dev server with hot reload

Environment Variables

| Variable | Description | Default | |----------|-------------|---------| | API_URL | AgentPlanner API URL | http://localhost:3000 | | USER_API_TOKEN | API token (required) | — | | MCP_TRANSPORT | stdio or http | stdio | | PORT | HTTP mode port | 3100 | | NODE_ENV | Environment | production |

License

MIT License - see LICENSE for details.

Support

• GitHub Issues
• CHANGELOG.md for version history
• PDR.md for technical design