@intentqa/mcp

A Model Context Protocol server over the intentqa CLI. JSON-RPC 2.0 on stdio.

{
  "mcpServers": {
    "intentqa": {
      "command": "npx",
      "args": ["-y", "@intentqa/mcp"]
    }
  }
}

intentqa init writes that file for you. To check the install without standing up a client:

npx @intentqa/mcp --tools

Tools

| Tool | Writes | | |---|---|---| | intentqa_context | | The catalog, controls, config, features, workflow, refusals and exit codes — one call. Start here. | | intentqa_catalog | | The vocabulary. view: json \| prompt \| snippets. | | intentqa_explain | | One entry: params, types, enum values, risk, an example line. | | intentqa_validate_feature | | Resolve draft Gherkin passed as text. Writes nothing. | | intentqa_controls | | The control inventory; with check and a url, what the app is missing. | | intentqa_lint | | Hard sleeps, focused tests, stale specs, raw locators. | | intentqa_audit | | Vocabulary size and reuse against the budget. | | intentqa_check | | The full pre-handoff gate. What CI runs. | | intentqa_trace | | Requirement-to-test coverage from @req: tags. | | intentqa_triage | | Classify failures in a Playwright JSON report. | | intentqa_generate | yes | Compile features to specs. check: true is read-only. | | intentqa_propose | yes | Queue a vocabulary gap for a human to turn into a step. | | intentqa_rules | | Workflow, refusals, exit codes. Works with no config present. |

Every tool takes an optional cwd, so one server can serve several projects. All but the two marked report readOnlyHint: true — a client can auto-approve them and save the prompt for the two that write.

The model still only picks catalog ids. It cannot invent a step.

What is deliberately absent

No locator repair. There is no tool that asks a model to fix or invent a selector. Run intentqa_controls with check and a url to get the list of missing data-testid attributes, then add them to the application. That is J7, and it is why a generated spec is worth reviewing.

No test runner. intentqa run launches browsers and can take minutes; holding a JSON-RPC call open for that is a bad trade. Run it in a shell and bring the report back through intentqa_triage.

Notes for anyone adding a tool

• Run the handler inside runCaptured. stdout on this transport carries JSON-RPC only, and every intentqa command prints for a human. One stray ui.success() corrupts the stream and disconnects the client.
• Set readOnly honestly. Clients auto-approve read-only tools.
• No MCP SDK dependency: the protocol is ~150 lines in src/protocol.ts, for the same reason the planner adapter is optional — nothing here should put a network install between the tool and a locked-down CI runner.

test/protocol.test.ts covers the handshake and framing against handleMessage directly; test/tools.test.ts drives the real tool set through tools/call against a project on disk.