ATP Librarian MCP Server
This server does not currently have a deployable release on Glama. Follow the repository's setup instructions to install and configure it for your MCP client, including any required credentials and local dependencies.
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.
Related MCP server: Agent Harness Kit
Local install
pip install -e .Run:
python3 main.pyConfigure your MCP client to call that command, and pass
plan_pathon 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-librarianEnvironment 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_pathdirectly; 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 aroundnode_id.
Resource
atp://status/summary: Plaintext dashboard with project status, live claims, and ready tasks for the default plan (ATP_FILEor.atp.jsonin 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 optionalcontext,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 byatp_decompose_taskonce allscope_childrenfinish.atp_read_graph(..., view_mode="full")includes runtimemeta.graph_versionin 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.
This server cannot be deployed
Maintenance
Related MCP Connectors
Shared task layer for AI agents: tasks, handoffs and decisions that persist across tools.
Task tracking built for coding agents. Work is leased, so two agents never take the same SubTask.
Task manager for AI agents. Orchestrate agents to achieve large goals in unified control plane.
Project management MCP for AI agents with safe task reads and writes.
Related MCP Servers
- AlicenseNot gradedqualityDmaintenanceEnforces that tasks are created with criteria, claimed by one agent with evidence, and verified by a different agent, preventing self-verification and ensuring independent validation.6 npmMIT
- AlicenseNot gradedqualityAmaintenanceEnables AI coding providers to coordinate structured multi-agent workflows with documents, tracked tasks, handoffs, and health evidence. It owns task claims, action logs, and final health results while leaving specification files on the filesystem.29 npmMIT
- AlicenseNot gradedqualityAmaintenanceEnables structured multi-agent workflows by managing task claims, action logs, handoffs, and health-gated completion while keeping specifications as filesystem Markdown artifacts.29 npmMIT
- AlicenseNot gradedqualityAmaintenanceEnables AI coding providers to run structured multi-agent workflows by claiming tasks, logging actions, managing handoffs, and enforcing server-side health gates across explorer, consultant, builder, and reviewer roles. Specifications stay as Markdown on disk while tracked task state, acceptance evidence, and completion checks are owned by the server.29 npmMIT