`bazel_set_workspace_path` means the model can point the server at whichever checkout the question is about, so one configuration entry covers every repository on the machine rather than one per project. `additionalArgs` keeps the escape hatch open — `--test_output=all` on a failing test gets you the output that the default summary hides. Set `--log_path` if your client swallows console logs, which is the usual reason a failure looks like silence.
A local MCP server that wraps the Bazel command line. It exists for environments where the agent has no usable shell, or where the shell it does have has an environment misconfigured enough that Bazel will not run in it.
- Targets built and tests run — `bazel_build_target`, `bazel_test_target`
- The dependency graph queried for targets matching a pattern — `bazel_query_target`
- Every target in the workspace listed, taking a path where `//` means all of them — `bazel_list_targets`
- External dependencies fetched — `bazel_fetch_dependencies`
- The workspace path changed at runtime, so one running server can move between checkouts — `bazel_set_workspace_path`
- An optional `additionalArgs` on every tool except the workspace switch, for flags like `--verbose_failures` or `--test_output=all`
No account and no key — it drives your local Bazel. The npm package name is `@nacgarg/bazel-mcp-server` (0.1.0) and it runs straight from GitHub with `npx -y github:nacgarg/bazel-mcp-server`. You usually do not need to configure anything: the Bazel binary is picked up automatically and the workspace can be set at runtime by the model. Where that fails, `--bazel_path` and `--workspace_path` take absolute paths, and each has an environment-variable equivalent in `MCP_BAZEL_PATH` and `MCP_WORKSPACE_PATH`, with a config file as the third fallback. Because it runs locally, it will not work through Cursor's Remote SSH sessions.
One command — npx -y github:nacgarg/bazel-mcp-server