Coverage MCP

Use Coverage MCP when you want a coding agent to run approved tests, remember coverage across sessions and Git worktrees, explain what changed, and point to the next useful test without pasting a large coverage report into the model's context.

Coverage MCP turns a project-generated coverage report into an immutable snapshot. An agent can then ask focused questions such as:

• Which changed lines are still untested?

• Did this worktree improve or regress coverage?

• What are the highest-priority gaps?

• What did one smaller test case add to a fixed full-suite baseline?

• Which named tests cover exactly the same regions and deserve human review?

The normal user experience is the testing plugin from codegen-marketplace. The plugin gives the agent a safe coverage-review workflow and starts the native Coverage MCP runtime automatically.

Safety note: Coverage MCP can run commands in your repository with your user account's permissions. A new command must be shown to you and explicitly approved before it can be registered or run.

Install for Codex

Do not clone this development repository for normal use. Install the published testing plugin instead:

codex plugin marketplace add appunni-m/codegen-marketplace
codex plugin add testing@codegen-marketplace

Start a new Codex task in the repository you want to review. The marketplace plugin downloads and verifies a compatible Coverage MCP release on supported macOS, Linux, and WSL hosts. If a prebuilt release is unavailable and Cargo is already installed, the plugin can build the exact published crate as a fallback.

For another MCP host, native Windows, or a deliberate manual installation, use the package published on crates.io:

cargo install coverage-mcp --locked
coverage-mcp --version

Manual Cargo installation requires Rust 1.85 or newer. See the installation and first-review guide for platform details, upgrades, and removal.

Related MCP server: SNAP MCP Server

Get your first coverage review

Open a new Codex task in your project and ask:

Use the coverage-review workflow. Inspect this repository's current coverage and show me the three highest-priority gaps. Reuse a fresh approved run if one exists. If a coverage command needs approval, show me the exact command, working directory, shell, report path, format, and suite before running it.

If the project has not been set up before, the agent will inspect its existing test tooling and propose a coverage command. Review that command and its report path before approving it. A successful first run reports:

• the command's terminal state and coverage-artifact status;

• an immutable snapshot ID;

• available line, branch, function, and region metrics;

• a bounded list of coverage gaps and one suggested next action.

After the daemon starts, the local dashboard is available at http://127.0.0.1:59471/. It shows stored snapshots, comparisons, source gaps, test runs, and retention status. You do not need to start the dashboard or daemon separately when using the Codex plugin.

How the workflow fits together

your approved test command
          │
          ▼
project coverage report ──► immutable snapshot ──► bounded agent review
                                      │
                                      └──────────► local dashboard and history

Coverage MCP does not generate language-specific coverage by itself. Your project's existing test tool produces LLVM JSON, LCOV, coverage.py JSON, Cobertura XML, JaCoCo XML, Istanbul JSON, or a Go cover profile. Coverage MCP validates and stores that evidence, preserves its Git and run provenance, and returns only the detail needed for the current question.

The system is local-first:

• the service binds only to loopback;

• each Git repository keeps its own coverage history;

• linked worktrees share the same repository history;

• approved commands and completed snapshots are durable;

• compact responses keep large reports out of the agent context.

Choose the guide you need

The MCP server is self-describing: its initialize and tools/list responses contain the live tool schemas and workflow constraints. The documentation is a human-readable guide; the executable response is authoritative for the installed version.

Status and compatibility

Coverage MCP is pre-1.0 software. The current source contract is schema revision 16 with eight agent-facing tools. Before 1.0, a minor release may make a breaking public-contract change; the marketplace pins a compatible runtime for each plugin release.

The source checkout pins Rust 1.85.1 and supports a manual Cargo install with Rust 1.85 or newer. Marketplace prebuilt releases target ARM64 and x86-64 macOS and Linux, including WSL. Native Windows currently uses the published Cargo package and an explicit coverage-mcp.exe connect client configuration.

The repository's required gate checks formatting, Clippy, tests, rustdoc, a bundled release build, and 100% measured line, function, and region coverage for the configured Rust library/runtime target set. That gate is a contributor policy, not a claim that every possible deployment or input is defect-free.

Security, support, and license

Coverage reports, repositories, retained logs, and approved commands are local inputs and may contain sensitive information. Keep the service on loopback and do not commit .coverage-mcp/ databases. See the security policy for the trust boundary and private vulnerability-reporting route.

Use the support guide for sanitized bug reports and feature requests. Coverage MCP is maintained on a best-effort basis without a commercial support SLA.

Coverage MCP is released under the MIT License.