MCP integration
A test management MCP server for your coding assistant
Your coding assistant writes the code. With TestPlum's MCP server, it can also read the requirement, write and link the tests, sort the failures and tell you whether the release is ready, from evidence instead of a guess.
TestPlum's MCP server connects AI coding assistants such as Claude Code, Cursor and VS Code to TestPlum's REST API with an organization API key. It has 17 tools for requirements, tests, runs, failures, flaky tests and release readiness, and it cannot delete anything.
What MCP is
The Model Context Protocol is an open-source standard for connecting AI applications to outside systems. A client such as Claude Code, Cursor or VS Code connects to an MCP server and can then use that server's tools.
TestPlum's MCP server runs on your machine, next to your assistant. It talks to TestPlum's REST API with an organization API key and can do exactly what that key's permissions allow, nothing more.
What you can ask
- Read requirement MOB-R1 and write tests for each acceptance criterion; put them in the Onboarding folder.
- Which requirements in MOB have no tests yet?
- I just wrote tests/checkout.spec.ts. Record each one as a TestPlum test with its automation ID.
- Are we ready to release Mobile Banking? What is blocking it?
- What failed in MOB-RUN6 and why?
- Which tests are flaky? Which should I run first today?
The tools
Items are named by key: MOB-C12 for a test, MOB-R3 for a requirement, MOB-RUN6 for a run. A key carries its project.
| Area | Tools |
|---|---|
| Projects | list_projects, get_project_structure |
| Tests | search_test_cases, get_test_case, create_test_case, update_test_case |
| Requirements | list_requirements, get_requirement, link_cases_to_requirement |
| Runs and results | list_runs, create_run, record_result |
| Failures | triage_run, get_run_failures |
| Risk and flakiness | get_risky_tests, get_flaky_tests |
| Release | get_release_readiness |
How to get it
The MCP server is not yet published as a public package, so there is no public install command to copy. It is available to workspaces on request. Once you have it, setup is two steps:
- Create an API key. In TestPlum, open API & integrations and create an organization API key. Treat it like a password.
- Add the server to your assistant. Give it TESTPLUM_URL (your TestPlum address) and TESTPLUM_API_KEY in environment variables, never in a file you commit. TESTPLUM_PROJECT is optional and sets the default project.
Guardrails
- Nothing deletes. None of the tools deletes data. Every change is versioned or append-only in TestPlum.
- The key sets the limits. The server can do only what the API key's permissions allow.
- AI-written tests count partly. Tests created through the MCP server are marked as AI-written. Their passes count only partly in the release verdict until a signed-in person reviews them, and an API key cannot do that review. An edit needs a new review.
- Fixes need a person. Fixes proposed to automated tests still need a person's approval, and TestPlum never writes to your repository.
A typical session, from requirement to verdict
- Read the requirement. The assistant calls get_requirement and reads the acceptance criteria and the tests that already verify it.
- Write and link tests. It writes the automated tests in your repository, then records each one in TestPlum with create_test_case, using the automation ID your results will carry, and links them with link_cases_to_requirement.
- Run them. Your pipeline runs the tests and sends the results, as on the GitHub Actions or GitLab CI pages. Results land on the tests the assistant created, matched by automation ID.
- Read the failures. If something fails, triage_run returns the failures grouped by cause with reasons, and get_run_failures gives the errors and step outcomes for each test.
- Ask for the verdict. get_release_readiness returns the verdict with each check and its evidence, the trusted pass rate and the weakest passes. A person still reviews the AI-written tests before their passes count in full.
MCP server or CI key?
They do different jobs. A CI key lives in your pipeline and can only send results, run the gate and read the status of one project. An organization API key powers the MCP server and acts as the person who created it, with their permissions across the organization.
So choose whose key the assistant uses with care: it can do whatever that person can do, and nothing more.
Why give an assistant test context
An assistant that can read acceptance criteria writes tests for what was asked, not just for what the code does. One that can read the readiness verdict answers "can we ship?" with the checks and their evidence, including the weakest passes and why they count partly. And one that can triage a run sees failures already grouped by cause and sorted as product bug, broken test, environment or flaky, with the reasons.
Frequently asked questions
Which assistants work with the TestPlum MCP server?
Any MCP client, such as Claude Code, Cursor, VS Code Copilot or Windsurf.
Is the MCP server on npm?
Not yet. It is not published as a public package; it is available to workspaces on request.
Can the assistant delete my tests?
No. None of the MCP tools deletes data, and every change is versioned or append-only.
Can the assistant mark its own tests as reviewed?
No. Reviews need a signed-in person, and API keys cannot review. Until then, passes of AI-written tests count only partly in the verdict.
Does it need a CI key or an API key?
An organization API key from API & integrations. CI keys are only for sending results and reading the status from a pipeline.
Where does the server run?
On your machine, next to your assistant. It calls TestPlum's REST API over HTTPS with your key.
Sources
- Model Context Protocol: What is MCP?, checked November 19, 2026
Claude Code, Cursor, VS Code and Windsurf are trademarks of their owners. TestPlum is not affiliated with or endorsed by them.