QE Craft MCP (@qe-craft/mcp)

Guided QE in your copilot — phased playbooks for people who are not full-time QEs. No API key. Inference stays in Cursor; MCP returns the director's script and optional artifact save.

Pair with seven Cursor skills (npx @qe-craft/mcp init): router + five mode skills + qe-automate for test writing.

Requirements

• Node.js 22+
• Cursor (stdio MCP)

Quick setup

1. MCP — ~/.cursor/mcp.json

{
  "mcpServers": {
    "qe-craft": {
      "command": "npx",
      "args": ["-y", "@qe-craft/mcp@latest"],
      "env": {
        "REPO_ROOT": "/absolute/path/to/your/target-repo"
      }
    }
  }
}

2. Skills

npx @qe-craft/mcp init
# team: npx @qe-craft/mcp init --project /absolute/path/to/your-repo
npx @qe-craft/mcp@latest init --force   # upgrade

Installs: qe-analysis, qe-refinement, qe-uat-gate, qe-repo-charter, qe-incident, qe-regression-impact, qe-automate.

3. Restart Cursor

| Variable | Purpose | |----------|---------| | REPO_ROOT | Repo for scan hints + docs/qe-analysis/ writes |

Trust-first default

• No LLM in the MCP server — qe_intel_* guided runs and optional validate/save only.
• No API key in MCP config — QE reasoning stays in your IDE agent (skills + phased playbook).
• Coach-first — better QE in chat by default; files only when you ask.

Data handling: See ../docs/data-handling.md.

This server does not call any external LLM API.

MCP tools

Primary (start here):

| Tool | Purpose | |------|---------| | qe_intel_refinement | Guided story grooming | | qe_intel_uat | Guided release gate | | qe_intel_repo_uat | Guided repo charter | | qe_intel_bug | Guided incident analysis | | qe_intel_regression | Guided retest scope | | qe_intel_review | Draft quality check |

No API key. Coach-tier success = better QE in chat, not a file on disk.

Architecture

Coach-tier UAT example: MCP returns the phased playbook; Cursor explores the repo and reasons in chat. Save to disk only when asked.

Example (UAT — coach)

Paste in chat:

Use qe-uat-gate. Call qe_intel_uat with:

feature: Release 2.4 — checkout promo ready for prod
release: { "rollback": "revert flag checkout.promo_v2", "monitoring": "promo_apply_failure_reason metric" }

Explore the repo, execute Phases A–D from the tool response, give GO/NO-GO in chat.
Do not save files unless I ask.

Agent should: one qe_intel_uat call → follow phases → optional qe_intel_review → save only on request.

Guided run (default)

| Step | Owner | Action | |------|--------|--------| | 1 | Cursor | Match intent → skill → qe_intel_<mode> with ticket/context (output_tier: coach) | | 2 | MCP | Return phased playbook (A–D), repo where to look hints under REPO_ROOT, input gaps if thin | | 3 | Cursor | Execute phases in chat — questions, risks, scenario table, GO/NO-GO / AC gaps | | 4 | Cursor (optional) | qe_intel_review on draft | | 5 | MCP (optional) | Phase E: validate + save only if user wants files |

See skills/shared/intel-run.md. Full 11-section / JSON contract: output_tier: full + artifact-run.md.

Output artifact table

| output_format | Files written (save_file=true) | MCP chat body | |-----------------|----------------------------------|---------------| | markdown (default) | docs/qe-analysis/qe-analysis-{MODE}-{slug}-{date}.md only | Full markdown or summary + Saved to: footer | | json | Same stem: .json (envelope) + .html — no .md | Short summary (mode, confidence, risks, scenario counts) + paths — not full HTML | | JSON parse/validate failure | Optional .raw.txt only if wired — not default | Error list + path to raw file when saved |

Collision suffix (-2, -3) applies to the stem before extension; sibling .json and .html share one stem.

Regenerate committed v2 samples after schema or renderer changes:

cd mcp && npm run build && node scripts/write-v2-samples.mjs

Artifact tools (Phase E only)

| Tool | Purpose | |------|---------| | qe_validate_report | JSON + evidence guards | | qe_save_report | .json + .html | | qe_save_markdown | .md | | qe_get_system_prompt | Debug / full-tier rules |

See skills/shared/artifact-run.md in this package.

Environment variables

| Variable | Required? | Purpose | |----------|-----------|---------| | REPO_ROOT | No | Absolute path to the repo where docs/qe-analysis/ should be written (defaults to MCP process cwd) |

There are no ANTHROPIC_MODEL, ANTHROPIC_MAX_TOKENS, or API-key variables for this server — model choice and token limits are entirely your IDE agent's provider. See .env.example.

Install from source (this repo)

cd mcp
npm install
npm run build
test -f dist/server.js && echo "Build OK"

Optional: REPO_ROOT=/absolute/path/to/target-repo so analyses save under that repo's docs/qe-analysis/ (defaults to process cwd).

Cursor MCP — local clone (~/.cursor/mcp.json)

Use absolute paths on your machine. No API keys.

{
  "mcpServers": {
    "qe-craft": {
      "command": "node",
      "args": [
        "/ABSOLUTE/PATH/qe-craft/mcp/dist/cli.js"
      ],
      "env": {
        "REPO_ROOT": "/ABSOLUTE/PATH/to/your/target-repo"
      }
    }
  }
}

Restart Cursor after saving.

Local dev (stdio):

cd mcp && npm run dev

Prompt hygiene

Prompts and bundled skills should stay vendor-neutral (no employer-specific product names). When coaching rules change, update src/intel/ and src/core/prompts/, sync skills under skills/, run npx @qe-craft/mcp init --force, and bump PROMPT_VERSION in src/core/constants.ts.

Publishing (maintainers)

1. cd mcp && npm run pack:check — dry-run pack after build + tests.
2. Confirm version in package.json and changelog intent.
3. npm login (org qe-craft for scoped publish).
4. npm publish --access public from mcp/ (prepublishOnly runs check + test).
5. Smoke: npx @qe-craft/mcp@latest init --dry-run and connect MCP in Cursor as qe-craft.

Develop

npm install && npm run check && npm test

Built by

An SDET actively piloting AI-assisted QE workflows in production — not a theoretical implementation.

Portfolio