tcm-mcp — TCM MCP Server (Epic 1: Test Case CRUD)

A stdio MCP server that gives AI agents (Torque, triage-e2e, Claude agents) a stable tool interface to read and write TCM test cases — without touching the database schema directly.

It is a thin client: every tool call proxies a TCM REST endpoint. ID resolution, validation, display_id generation, the in_cicd lock, and soft-delete scoping all happen inside TCM. Agents reference cases by display_id (e.g. APA-3); internal UUIDs are never exposed.

Full design: docs/features/mcp-e1-test-case-crud.md (in the main TCM repo).

Tools

Reads exclude trashed (soft-deleted) cases. Writes require the dry-run flow.

Project scoping. search_suite, list_suites, and list_test_cases scope by project. Pass a project_id (UUID) directly, or a project_name — the server resolves the name to an id via list_projects (case-insensitive exact match; an unknown name returns NOT_FOUND and an ambiguous one returns AMBIGUOUS with the candidate ids). Use list_projects first to discover ids. list_test_cases with no project returns cases across every project you can see.

Related MCP server: TestRail MCP Server

Requirements

• Node.js ≥ 18 (for npx and the global fetch).

• Git read access to JoinFullStackDev/tcm-mcp — the package is distributed by git URL, not published to npm. On headless hosts (OpenClaw/Torque) a git token must be present in the environment.

• That's it for the TCM URL: it defaults to production (https://tcm-ochre.vercel.app), so there's nothing to look up or set. You just need to authenticate (Quickstart).

Because it's distributed by git URL, npx clones the repo and builds from source on first run (via the package's prepare → tsc step), so the first launch is slower. Subsequent runs are cached.

Quickstart

Zero-config: the production TCM instance (https://tcm-ochre.vercel.app) is baked in as the default, so you never set TCM_BASE_URL. Point at a different instance only if you self-host (see Environment variables).

Prerequisite — git access. This package is fetched by git URL from a private repo, so the machine running it needs git read access (gh auth login, or a git token for headless hosts). Node ≥ 18 must be installed. On macOS, GUI-launched Claude Desktop may not see your shell PATH — if the server fails to start, use an absolute path to npx in the config (find it with which npx).

1. Register the server

Claude Code — one command, nothing to edit by hand:

claude mcp add tcm --scope user -- npx --yes github:JoinFullStackDev/tcm-mcp#v1.4.0 --stdio

--scope user makes it available in every project. (Drop it to scope to the current project; Claude Code writes the .mcp.json for you.)

Claude Desktop — no CLI, so add it to the config file once:

1. Settings → Developer → Edit Config — this creates and opens claude_desktop_config.json for you (no folder to make yourself):

   • macOS: ~/Library/Application Support/Claude/claude_desktop_config.json

   • Windows: %APPDATA%\Claude\claude_desktop_config.json

2. Add the tcm entry (merge into mcpServers if it already exists), then fully quit + reopen Claude Desktop:

{
  "mcpServers": {
    "tcm": {
      "command": "npx",
      "args": ["--yes", "github:JoinFullStackDev/tcm-mcp#v1.4.0", "--stdio"]
    }
  }
}

Pin to a tag (#v1.4.0), not a branch — a branch ref re-resolves on every launch and can trip the 30 s MCP startup timeout. No env block is needed.

2. Sign in

The server starts even before you've logged in — it just exposes a login tool. So the easiest way (works in Claude Desktop and Claude Code, no terminal):

Just ask Claude: "Log me into TCM."

Claude calls the login tool, a browser opens once for Google sign-in, and the server stores a session it then keeps refreshed. The other tools light up immediately after. (Playwright is auto-installed on first login — a one-time ~100 MB browser download into ~/.tcm-mcp.)

Prefer a terminal? Same thing, run once:

npx --yes github:JoinFullStackDev/tcm-mcp#v1.4.0 login

Details: Auto-refreshing login.

If you'd rather edit .mcp.json directly (Claude Code project or ~/.claude/.mcp.json), the minimal entry is just command + args as shown above. To use the legacy static-token mode instead of a login session (e.g. CI that already has a JWT), add an env block — note it expires ~1h and refreshing it needs a full client restart:

{
  "mcpServers": {
    "tcm": {
      "command": "npx",
      "args": ["--yes", "github:JoinFullStackDev/tcm-mcp#v1.4.0", "--stdio"],
      "env": { "TCM_USER_TOKEN": "${TCM_USER_TOKEN}" },
    },
  },
}

Auth modes

The server resolves its mode at startup. Precedence: CLUTCH_API_KEY → login session file → TCM_USER_TOKEN.

In refreshing mode the server auto-renews the access token before expiry and again on any 401 (retrying the request once), and persists the rotated refresh token back to the session file. In static and clutch modes a 401 is terminal (nothing to refresh).

In headless mode you must also set MCP_AGENT_USER_ID, or create/update will fail on the created_by/updated_by NOT NULL constraint. The server prints a startup warning if it's missing. When set, the server forwards it to TCM as an X-Agent-User-Id header (trusted only alongside a valid X-Clutch-Key), so each agent attributes its own writes; TCM falls back to its own MCP_AGENT_USER_ID env if the header is absent. (Requires TCM with the matching write-attribution support.)

Auto-refreshing login (recommended)

The login helper signs you into TCM in a browser once and writes a session file the server then uses to keep itself authenticated indefinitely — no ~1h token churn, no client restarts.

# no clone needed — runs straight from the git URL:
npx --yes github:JoinFullStackDev/tcm-mcp#v1.4.0 login

# ...or, from a local clone of this repo:
npm run login

• Opens a browser only if there's no valid saved session; later runs refresh silently (headless, no window).

• Playwright is installed for you on first login. It is deliberately not a server dependency (keeps npx <server> installs lean ~50 MB), so the login helper installs playwright + Chromium once into ~/.tcm-mcp (a ~100 MB one-time download) if they aren't already present. You do not need a separate "Playwright MCP" — the login is fully self-contained.

It writes ~/.tcm-mcp/session.json (mode 0600) containing the Supabase project URL, anon key (public), and the access + refresh tokens. From then on the MCP server (mode “refreshing token”) mints fresh access tokens on demand.

• Session file location: ~/.tcm-mcp/session.json, override with TCM_SESSION_FILE.

• Browser profile: ~/.tcm-mcp/browser, override with TCM_BROWSER_PROFILE.

• Security: the refresh token is a long-lived credential — the file is 0600 and must never be committed or shared. Supabase rotates the refresh token on every refresh; the server persists the new one atomically.

• When it expires: if the refresh token is ever revoked/expired, tool calls fail with a clear “run npm run login” message. Re-run the helper.

• Anon key capture: the helper sniffs the public apikey header from Supabase network traffic. If capture ever fails, set SUPABASE_ANON_KEY (safe to expose) and re-run.

Environment variables

The recommended credential is the login session file (npm run login), not TCM_USER_TOKEN — see Auto-refreshing login. TCM_USER_TOKEN remains for CI / scripts that already have a JWT.

The write safety flow (dry-run → approval → commit)

create_test_case and update_test_case are two-pass:

1. Call with dry_run: true first. The tool validates, resolves IDs, and returns a summary (create: the proposed case; update: a field-level diff + before/after steps). No write happens.

2. A human reviews and approves — Torque relays the summary to Slack via Clutch; Claude Code shows it inline in the chat.

3. Call again with dry_run: false (or omit dry_run) to commit.

The server does not technically enforce that a dry-run/approval happened before a commit (decided: PRD OQ-4 Option A) — it's a process convention. Don't call with dry_run: false without human approval.

Local development

git clone https://github.com/JoinFullStackDev/tcm-mcp && cd tcm-mcp
npm install                 # runs prepare → tsc → dist/
npm run build               # rebuild after changes

# run the stdio server directly (Ctrl-D / EOF to exit)
TCM_BASE_URL=https://your-tcm-instance.example.com \
TCM_USER_TOKEN=your-jwt \
node dist/index.js

npm run dev                 # same, via ts-node (no build step)

Startup logs (mode, base URL, "Ready") are written to stderr, so they don't interfere with the stdio MCP protocol on stdout.

Notes & caveats

• Audit logging (mcp_tool_calls, PRD Appendix C) requires migration 00042 applied to the TCM database. The log inserts are fire-and-forget and non-blocking — if the table is missing, tools still work; only the audit trail is skipped.

• Distribution is git-URL only (no npm publish). Pin a tag; ensure hosts have git access.

• The --stdio arg in the config is cosmetic — stdio is the only transport.