Data Workers on OpenCode

Install it in OpenCode

OpenCode reads MCP servers from an opencode.json file in your project root, under an mcp key, with each server given as a command array rather than a command plus args.

git clone https://github.com/DataWorkersProject/dataworkers-claw-community.git
cd dataworkers-claw-community
npm install --ignore-optional

{
  "mcp": {
    "dw-pipelines": {
      "type": "local",
      "command": ["/path/to/dataworkers-claw-community/start-agent.sh", "dw-pipelines"],
      "enabled": true
    },
    "dw-catalog": {
      "type": "local",
      "command": ["/path/to/dataworkers-claw-community/start-agent.sh", "dw-context-catalog"],
      "enabled": true
    }
  }
}

Read on 11 September 2026 from the repository README, which is the source of truth. Note the shape difference from Claude Code and Cursor: OpenCode takes one command array, not a command string plus an args array, and each server needs the type and enabled fields. The repository ships a working opencode.json, so cloning it and opening OpenCode inside it needs no configuration at all.

Source, licence and the full agent list: the repository.

What to ask it in the first fifteen minutes

All three are read-only and all three are gradeable, which is the point. Pick questions where you already know the answer, because a data agent that sounds right is worth nothing and one you have checked is worth something.

1. Ask it to explain a table or model nobody on your team fully understands, then compare its account with your best engineer's. This is the cheapest test of whether the context layer is real.
2. Point it at the last incident you resolved by hand and ask for a root cause. You already know the answer, so you can grade it.
3. Ask which tables in one schema have no owner and no description, then decide whether that list is worth acting on. It is usually longer than people expect.

Both halves of this pairing are open source and run on your own model key, so nothing on the read-only path requires a commercial relationship with a vendor, including us.

Where OpenCode stops and the platform starts

This is the honest part of the page, and it is the reason the rest of it is worth trusting. An MCP server inside a coding agent is a very good pair of read-only hands that exists only while you are at the keyboard. Four things it will not do:

• Run when OpenCode is closed. A pipeline that breaks at 03:00 is still broken at 09:00, and the agent learns about it when you do.
• Write to your warehouse. The free agents read, analyse and recommend. The write tools are in the paid tier, deliberately.
• Hold an approval gate. There is no queue where a human approves a change before it lands, because there is no change landing.
• Keep an audit trail anyone else can read. Your session is your session.

Those four are exactly what the Autonomous Data-Conductor adds: it runs the detect, diagnose, fix and verify loop continuously, holds every write behind a policy you set, and records what happened where your team and your auditor can see it. What that category has to do to qualify, stated generically enough to use on any vendor: the autonomous agentic data platform.

If read-only agents in your editor are all your team needs, the free core is the right answer and you should not buy anything from us. That is a real outcome and we would rather you reach it in an afternoon than after a procurement cycle.

Common questions

The same install, elsewhere

• Data Workers on Claude Code
• Data Workers on Cursor
• Data Workers on Codex CLI
• Data Workers on Gemini CLI

And what the full platform does on the warehouse underneath:

• Data Workers for Databricks
• Data Workers for Snowflake
• Data Workers for dbt
• Data Workers for BigQuery
• Data Workers for Microsoft Fabric
• Data Workers for Airflow and Dagster
• Data Workers for Fivetran and Airbyte
• Whether you should use us at all, including the cases where the answer is no: the evaluation guide.