ATP Librarian MCP Server

This server enforces the Agent Task Protocol (.atp.json) contract so worker agents can claim, finish, and decompose tasks without corrupting the dependency graph. It exposes MCP tools that wrap deterministic updates to the plan file on disk.

Quick start (no clone, uvx)

Add to your MCP client config and update the plan path:

[mcp_servers.atp]
command = "uvx"
args = ["--from", "git+https://github.com/Edgeworthless/atp-mcp-server", "atp-server"]
env = { "ATP_FILE" = "/path/to/your/.atp.json", "ATP_LEASE_SECONDS" = "600" }

uvx downloads/builds on the fly. Tools still require plan_path per call; ATP_FILE is just a default.

Local install

1. pip install -e .
2. Run: python3 main.py
3. Configure your MCP client to call that command, and pass plan_path on every tool call.

Docker (optional)

Build or pull, then mount your plan file:

docker build -t atp-librarian .
docker run --rm -v /path/to/.atp.json:/data/.atp.json -e ATP_FILE=/data/.atp.json atp-librarian

Environment knobs

• ATP_FILE: path to the ATP plan file (default .atp.json).
• ATP_LOCK_FILE: override the lock file path (defaults to <ATP_FILE>.lock).
• ATP_LEASE_SECONDS: duration before a claimed task is released (default 600s).
• Tools accept plan_path directly; the env vars mainly backstop the status resource or when clients omit the argument.

Available tools

• atp_claim_task(plan_path, agent_id): Assigns the highest-priority READY node, recovers zombies, and returns dependency context. Re-enters an existing claim for the same agent.
• atp_complete_task(plan_path, node_id, report, artifacts=[], status="DONE"): Marks a task DONE or FAILED, clears the lease, and unlocks children whose dependencies are satisfied.
• atp_decompose_task(plan_path, parent_id, subtasks): Converts a task into a SCOPE with a new subgraph. Subtasks must form a DAG; start nodes inherit the parent’s original dependencies. The parent closes automatically once all new tasks complete.
• atp_apply_future_patch(plan_path, expected_graph_version, patch, reason, actor_id): Applies bounded future-graph mutations for ATP v1.4 adaptive judge flows. Apply is rejected unless claimed-node count is zero, the graph version matches, terminal/claimed nodes remain untouched, and the candidate graph stays valid.
• atp_read_graph(plan_path, view_mode="full" | "local", node_id=None): Returns the full JSON graph or a neighborhood view around node_id.

Resource

• atp://status/summary: Plaintext dashboard with project status, live claims, and ready tasks for the default plan (ATP_FILE or .atp.json in the current working directory).

Graph contract

• Keep the plan as a valid task DAG: each node has title, instruction, dependencies, status (LOCKED|READY|CLAIMED|COMPLETED|FAILED), plus optional context, reasoning_effort (minimal|low|medium|high|xhigh), artifacts, report, timestamps.
• Node statuses: LOCKED (waiting), READY, CLAIMED, COMPLETED, FAILED. The server will auto-release CLAIMED tasks after the lease window and auto-complete SCOPE nodes produced by atp_decompose_task once all scope_children finish.
• atp_read_graph(..., view_mode="full") includes runtime meta.graph_version in its response payload for optimistic concurrency in adaptive judge flows.
• ATP v1.4 future-graph mutation is additive. Closed/superseded future nodes are tracked via optional metadata and are not claimable or counted as READY work.
• Dependency integrity is checked on every read/write; missing node references will raise errors instead of silently mutating the file.

Testing

• python3 -m compileall main.py

MCP client config example

Example JSON config (Codex-compatible):

{
  "mcpServers": {
    "atp-librarian": {
      "command": "python3",
      "args": ["/path/to/atp-mcp-server/main.py"],
      "cwd": "/path/to/atp-mcp-server",
      "env": {
        "ATP_FILE": "/path/to/your/project/.atp.json",
        "ATP_LEASE_SECONDS": "600"
      }
    }
  }
}

Tool calls should still provide plan_path (e.g., /path/to/your/project/.atp.json); ATP_FILE just sets a default for the status resource or when a client omits the argument.