MCP server for live coordination between AI coding agents. Exposes coordination tools over stdio MCP so both Claude Code and Codex can connect to the same project coordinator.
The server defines 7 coordination tools with clear, domain-specific purposes (task claiming, ownership release, contract publishing, progress updates, event subscription). All tools have descriptions (avg 95 chars) and input schemas with typed parameters. However, schemas lack depth: most parameters are strings without format constraints, enums, or length bounds. Output schemas are entirely undocumented, callers cannot predict response structure. Error handling is absent from descriptions; no guidance on retryability, recovery, or failure modes. Parameter descriptions are minimal (avg 45 chars) and lack actionable constraints. The tools follow a single-responsibility pattern well (claim vs release vs publish), but composition guidance is missing.
Acknowledge that an agent has incorporated events up to a sequence number.
Reserve a task and optionally declare file ownership. Returns the task event and any ownership conflicts.
Publish a versioned API contract (endpoint, schema, types). Other agents are notified of the new revision.
Report progress: files touched, commands run, blockers, next action.
Release exact file or directory paths previously owned by this agent.
Get a bounded summary of current coordination state for a reconnecting agent: pending tasks, active contracts, file ownership, unacked events.
Read events after a cursor (sequence number). Returns new events and the current high watermark for reconnection.
Output schemas completely undocumented. No tool describes what fields are returned, their types, or structure. LLMs cannot plan downstream calls or extract required data.
Parameter descriptions lack actionable constraints. 'agent' is described as 'Agent claiming the task (e.g. claude-code, codex)', examples are not constraints. Should use enum or pattern. 'taskId' has no format guidance. 'ownPaths' lacks validation rules (absolute vs relative, max count).
No error handling guidance in any tool description. No indication of retryability, failure modes, or recovery steps. E.g., claim_task could fail if task is already claimed, but description does not explain what to do.
| Scored | Grade | Overall | Spec posture | Rubric |
|---|---|---|---|---|
| 2026-09-23 | D | 58 | 2026-07-28+ | v2 |
Parameter 'agent' appears in all tools but is never formally constrained. Should be an enum of valid agent IDs (claude-code, codex, cursor, etc.) to prevent hallucinated agent names.
publish_contract 'format' parameter has no enum. Description says 'Format hint: typescript, openapi-yaml, sql, json-schema' but these are examples, not constraints. LLM may pass invalid formats.