Tools

The server intentionally exposes exactly three tools. Task status and results, when negotiated, use MCP protocol methods rather than additional tools.

bazel.run

Run an allowed Bazel command and receive a bounded result.

{
  "workspace": "/src/project",
  "command": "test",
  "args": ["//services/...", "--test_tag_filters=-integration"]
}

Supported command classes include build, test, coverage, query, cquery, aquery, and selected informational commands. The exact policy is configurable. Request arguments remain an array and are never concatenated into a shell command.

The result includes the invocation lifecycle state, command outcome, a bounded headline and diagnostics, and an invocation_id when evidence can be inspected.

bazel.inspect

Read a narrow view from a retained invocation. Available views include:

• summary and diagnostics;
• test results and copied failed-test logs;
• coverage and artifacts;
• query results;
• normalized logs.

{
  "invocation_id": "019f...",
  "view": "test_log",
  "limit": 20
}

Views are filtered, byte-bounded, and paginated. When a response is truncated, pass its opaque next_cursor to the next inspection request without modifying it.

bazel.cancel

Cancel a queued or running invocation.

{
  "invocation_id": "019f..."
}

Cancellation uses the same durable invocation ID as execution and inspection. The server coordinates interrupt and termination grace periods rather than leaving process cleanup to the agent.

Long-running calls

When the client and server negotiate compatible MCP task support, bazel.run can return a durable task handle. Task polling and result retrieval are MCP protocol methods; the three-tool inventory does not change.

See long-running builds for execution modes.