Skip to content
Model Context Protocol

Replication you can
talk to.

The TwinCommit MCP server connects your replication pipelines to Claude, GitHub Copilot, Cursor and any MCP-compatible assistant. Install it with one command, then check lag, diagnose errors, re-snapshot tables and verify data by asking.

Assistant · twincommit MCP connected10 tools
Why is lag on db2-prod → databricks rising?
get_lag pipeline="db2-prod-databricks" window="1h"
get_status pipeline="db2-prod-databricks"
Lag climbed from 0.4s to 38s starting 14:02. There's one open transaction,txn 991204, with 2.1M rows on BILLING.LEDGER. It's spilling to disk, so it will apply as one unit at commit. Nothing is broken.
Likely cause: the month-end batch job. Want me to raise max_txn_memory_bytes to 1 GiB for this pipeline?
No, just tell me when it's caught up.
Watching twincommit_caught_up…

Super-easy install

One command. Every client.

No servers to deploy and no API keys to mint. The MCP server runs locally next to your assistant and talks to the TwinCommit status endpoint you already have.

  1. 1

    Install

    Add the server to your MCP client with one command.

  2. 2

    Connect

    Point TWINCOMMIT_URL at a running pipeline status endpoint (default http://localhost:9108).

  3. 3

    Ask

    Ask your assistant about lag, errors, broken tables or data verification in plain English.

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

Tools

10 tools for day-two operations

Every tool maps to a TwinCommit capability you can also script with the CLI.

list_pipelines

Every pipeline with its source, target, state and current lag.

get_status

Checkpoint position, open transactions, broken tables and last error.

get_lag

Lag history for a pipeline over a time window, from Prometheus or /status.

start_pipeline

Start or resume a pipeline from its last committed checkpoint.

stop_pipeline

Drain in-flight transactions and stop cleanly.

resnapshot_table

Re-snapshot one broken table while the rest keep streaming.

verify_tables

Compare source and target row counts and checksums.

explain_error

Turn a log or apply error into a cause and a concrete fix.

check_source

Run source prerequisite checks and return the exact fix commands.

tail_events

Stream recent change events for a table (values redacted by default).

Try asking

Things on-call engineers actually ask

“Which pipelines are behind by more than 10 seconds?”
“Why did db2-prod → databricks stop at 03:12?”
“Re-snapshot SALES.ORDERS on the postgres pipeline.”
“Verify row counts for every BILLING table.”
“Is my new Oracle source ready? Tell me what to fix.”
“Summarize replication health for the last 24 hours.”

Read-only by default

Write tools need TWINCOMMIT_MCP_ALLOW_WRITES=true. Nothing changes unless you say so.

Row values stay private

Tools return metadata only. Row values are redacted unless you explicitly opt in.

Every call audited

Each tool call is logged with its arguments, result status and the client that made it.

MCP server FAQ

What is the TwinCommit MCP server?

+

It is a Model Context Protocol server that exposes TwinCommit pipelines as tools to AI assistants such as Claude, GitHub Copilot and Cursor, so you can monitor and operate replication in plain English.

How do I install the TwinCommit MCP server?

+

Run claude mcp add twincommit -- npx -y @twincommit/mcp for Claude Code, or add the same npx command to the MCP configuration of Claude Desktop, VS Code or Cursor.

Can the assistant change my pipelines?

+

Only if you allow it. The server is read-only by default. Start, stop and re-snapshot tools need TWINCOMMIT_MCP_ALLOW_WRITES=true, and every call is written to an audit log.

Does row data get sent to the AI model?

+

Not by default. Tools return metadata such as positions, counts, lag and errors. Row values are redacted unless you explicitly enable TWINCOMMIT_MCP_ALLOW_ROW_VALUES.

Give your assistant a replication superpower.

Install the MCP server, point it at a pipeline and ask your first question in under a minute.