@deitum/testops-mcp

An MCP server for Allure TestOps: search, read and write test cases, author a whole suite in Markdown and import it in one call, manage launches, and read project statistics — all locked to a single project.

npx @deitum/testops-mcp

It speaks MCP over stdio and is configured entirely through environment variables.

The server is deliberately single-project: TESTOPS_PROJECT_ID is the only source of the target project, it is never part of a tool's input, and the model cannot reach another project by asking. Pointing it somewhere else means changing the configuration and restarting.

Setting it up in a client

Claude Desktop (claude_desktop_config.json), Cursor, Windsurf and anything else that reads the same shape:

{
  "mcpServers": {
    "testops": {
      "command": "npx",
      "args": ["-y", "@deitum/testops-mcp"],
      "env": {
        "TESTOPS_BASE_URL": "https://testops.example.com",
        "TESTOPS_API_KEY": "your-api-token",
        "TESTOPS_PROJECT_ID": "23",
        "TESTOPS_IMPORT_ROOTS": "/Users/you/test-cases"
      }
    }
  }
}

Claude Code:

claude mcp add testops \
  --env TESTOPS_BASE_URL=https://testops.example.com \
  --env TESTOPS_API_KEY=your-api-token \
  --env TESTOPS_PROJECT_ID=23 \
  -- npx -y @deitum/testops-mcp

VS Code (.mcp.json / .vscode/mcp.json), where the token is prompted for rather than written down:

{
  "inputs": [
    {
      "id": "testops-token",
      "type": "promptString",
      "description": "Allure TestOps API token",
      "password": true
    }
  ],
  "servers": {
    "testops": {
      "command": "npx",
      "args": ["-y", "@deitum/testops-mcp"],
      "env": {
        "TESTOPS_BASE_URL": "https://testops.example.com",
        "TESTOPS_API_KEY": "${input:testops-token}",
        "TESTOPS_PROJECT_ID": "23"
      }
    }
  }
}

OpenCode (~/.config/opencode/opencode.json, or %USERPROFILE%\.config\opencode\opencode.json on Windows). Note that permission sits beside mcp, not inside it:

{
  "$schema": "https://opencode.ai/config.json",
  "mcp": {
    "testops": {
      "type": "local",
      "enabled": true,
      "command": ["npx", "-y", "@deitum/testops-mcp"],
      "environment": {
        "TESTOPS_BASE_URL": "https://testops.example.com",
        "TESTOPS_API_KEY": "your-api-token",
        "TESTOPS_PROJECT_ID": "23",
        "TESTOPS_IMPORT_ROOTS": "/Users/you/test-cases"
      }
    }
  },
  "permission": {
    "testops_testcase_delete": "ask"
  }
}

The API token comes from TestOps itself: your avatar → API Tokens → + Token.

Related MCP server: allure-testops-mcp

Configuration

TESTOPS_BASE_URL must be https outside localhost and must carry no credentials, query, fragment or path. TESTOPS_JIRA_BASE_URL follows the same rules but may carry a path, since Jira is often published under one (https://jira.example.com/jira).

Issue links

Issue keys — from testcase_add_issue_links, or from a Ссылки / Issues section in an imported file — become external links of the form <TESTOPS_JIRA_BASE_URL>/browse/KEY. They are external links, not integrated TestOps Issues.

With no Jira root configured the server does not guess a host: testcase_add_issue_links fails before it writes anything, and an import file containing issue keys is rejected during preflight, so no case is created.

TLS

An instance behind a certificate Node does not trust fails with SELF_SIGNED_CERT_IN_CHAIN. Prefer NODE_EXTRA_CA_CERTS=/path/to/ca.pem, which trusts that one chain. TESTOPS_SSL_VERIFY=false is the blunt alternative — it turns verification off for every TestOps request the process makes — and is worth it only for an internal instance whose certificate you have already accepted by other means.

Safety model

• One project. Every project-scoped request carries the configured ID, injected by the client rather than taken from the model. Before a by-ID write, the entity is loaded and refused if it belongs to another project.

• Deletion means Trash. testcase_delete moves exactly one case — named exactly as testcase_get returned it — to Trash (force=false). There is no bulk and no permanent delete, and deletion is never used as an import rollback.

• Writes are never replayed on their own. After an expired token only reads are retried. A write whose outcome is unknown is reported as unknown: the model is told to reconcile and wait, not to try again.

• Import is preflighted, then non-transactional. Nothing is created until the whole file and the project's custom-field catalog have been validated. After that, cases are created one at a time: if one fails, the earlier ones remain, and the report says exactly which are created, partiallyCreated, failedCase and notAttempted.

Authoring and importing Markdown

A whole suite can be written as one Markdown file and imported in a single call. The format carries a description, preconditions, nested steps with individual expected results, tags, issue keys and custom fields addressed by their human-readable names — the server resolves those against the project's catalog.

The intended flow is: ask the model for the file, read it yourself, then ask for the import. The rules do not need to be pasted into the conversation — testcase_markdown_guide hands the model the canonical format and a worked example, and the model is instructed to call it before writing or repairing a file. A file that is already written goes straight to testcase_import_markdown.

The file must be an absolute path inside one of TESTOPS_IMPORT_ROOTS. For a human reference: the format and the example.

Tools

Test cases:

• testcase_search — one page of active cases by partial name and exact workflow status.

• testcase_get — one test case by ID.

• testcase_steps — a case's scenario tree, its steps and expected results.

• testcase_history — a case's execution history.

• testcase_search_by_custom_field — one server-side search by exact custom-field name and value.

• testcase_custom_fields — the project's custom fields and their allowed values.

• testcase_markdown_guide — the Markdown rules and the example, for the model to write against.

• testcase_import_markdown — validate a Markdown file and import every case in it.

• testcase_create — create one simple test case.

• testcase_add_steps — append steps, each with its own expected results.

• testcase_set_tags — replace a case's whole tag set.

• testcase_add_issue_links — add Jira keys as external links on one active case.

• testcase_delete — move exactly one confirmed case to Trash.

• testcase_trash_search — find deleted cases by ID or by name.

• testcase_restore — restore exactly one case chosen by ID and name.

Launches:

• launch_list — one page of launches.

• launch_get — one launch by ID.

• launch_create — create a launch.

• launch_update — rename a launch, given its exact current name.

• launch_statistic — a launch's result counts by status.

Project:

• statistics_overall — launch and test-case counts, and the automation percentage.

Search tools return one requested page with its metadata and accept at most 50 items, so a large project cannot land in the model's context whole. The AQL behind a search is built by the server: no tool accepts a raw query.

Troubleshooting

Using it as a library

The stdio binary is the point, but the server is exported too — for a custom transport or a test harness:

import { createTestOpsServer, loadConfig, TestOpsClient } from '@deitum/testops-mcp';

const config = loadConfig();
const client = new TestOpsClient(config.baseUrl, config.apiKey, config.projectId, {
  sslVerify: config.sslVerify,
});
const server = createTestOpsServer(client, { allowedMarkdownRoots: config.importRoots });
await server.connect(myTransport);

Development

npm install
npm run verify        # lint, format, types, tests, build, packaging
npm test              # vitest, no network — every TestOps call goes to a local fixture
npm run inspector     # build, then drive it with the MCP inspector

See CONTRIBUTING.md.

Licence

MIT.