Launch Post Draft

Publication gate: post this only after the PyPI package resolves, the GitHub repository has its launch metadata set, and Discussions is enabled.

Title Options

• Completed is not accepted: Coordination Memory for multi-agent coding
• A local MCP coordination layer for parallel coding agents
• Stop coding agents from overwriting each other or self-certifying done

Long Form

When multiple coding agents work on the same project, two things go wrong fast: they overwrite each other, and "the agent says it is done" gets confused with "the work was accepted."

Coordination Memory MCP is a local-first MCP server for that coordination layer. It records assignment ownership, run liveness, human interventions, handoff evidence, and integrator review decisions in an append-only SQLite store.

The key rule is simple: completed is not accepted.

An agent can submit a handoff with evidence. Only an integrator can promote that handoff into the accepted ledger. For goal-level work, acceptance contracts make self-certified done structurally insufficient: criteria are sealed, probes are reported as evidence, and an independent acceptor signs off.

It is intentionally boring infrastructure:

• local SQLite and stdio MCP
• no deployment
• no secret reading
• no shell execution
• optimistic concurrency on mutating writes
• exported accepted projection for durable audit
• a local dashboard for assignments, runs, reviews, and contracts

Install:

uv tool install coordination-memory-mcp
comem --help

Run:

COORDINATION_MEMORY_DB=/absolute/path/coordination.sqlite3 comem serve

Set up agent onboarding files:

comem init

Open the local dashboard:

COORDINATION_MEMORY_DB=/absolute/path/coordination.sqlite3 \
  comem dashboard --host 127.0.0.1 --port 8765

Good fits:

• multiple Codex, Claude Code, Cursor, or other agents working from separate worktrees
• teams that need review before an agent's work counts as accepted
• maintainers who want a replayable audit trail of coordination decisions
• local-first users who do not want their coordination layer reading secrets or executing commands

Not a fit:

• generic long-term memory
• RAG
• secret storage
• cloud orchestration

Repository: https://github.com/yanqiw/comem

Short Form

I released Coordination Memory MCP: a local SQLite-backed MCP server for multi-agent coding coordination.

The key idea: completed is not accepted.

Agents can claim leased assignments and submit handoff evidence. Only an integrator can accept work into the ledger. Acceptance contracts add independent goal-level sign-off.

Repo: https://github.com/yanqiw/comem

Outreach Checklist

• GitHub repository release notes for v0.2.4.
• MCP Registry once PyPI is live and server.json is publishable.
• Glama MCP server directory.
• awesome-mcp-servers pull request.
• MCP community channels with a feedback request focused on multi-agent coding coordination.
• Hacker News "Show HN" only after install works from PyPI.
• X/LinkedIn short form with the dashboard screenshot.