Managing CDC Pipelines with MCP: Replication You Can Talk To

The TwinCommit MCP server lets an AI assistant operate your replication pipelines. Install it with one command, and Claude, GitHub Copilot or Cursor can check lag, explain errors, verify tables and, if you allow it, restart pipelines or re-snapshot a table.

One-line install

claude mcp add twincommit -- npx -y @twincommit/mcp

That’s it. The server runs locally and talks to the TwinCommit /status endpoint (default http://localhost:9108).

What on-call looks like now

You: Why is lag on db2-prod → databricks rising?

Assistant: One open transaction with 2.1M rows on BILLING.LEDGER is spilling to disk. It will apply atomically at commit. Nothing is broken.

Behind the scenes the assistant called get_lag and get_status, the same data you would piece together from Grafana and logs.

Safe by default

• Read-only unless TWINCOMMIT_MCP_ALLOW_WRITES=true
• Row values redacted unless explicitly allowed
• Every call audit-logged

Replication is critical infrastructure. Assistants should make it easier to understand, not easier to break.