qai-consultant-mcp

Listed on the official MCP registry (io.github.gvasile29/qai-consultant-mcp), Glama, and Awesome MCP Servers.

A local, fully keyless MCP server: standards-grounded QA knowledge retrieval (OWASP, IEEE, ISO, A-SPICE, EU AI Act, plus testing methodologies and audit frameworks), deterministic QA effort estimation, QA document quality review, test-results health analysis, and QA process maturity assessment — callable from Claude Code, Claude Desktop, or any other MCP client that runs local stdio servers.

No API keys, no Pinecone, no cloud LLM calls. It runs a local embedding index over a self-authored QA knowledge base (Markdown only — the ISTQB syllabus and OWASP guide PDFs used by the web app are not bundled, for licensing reasons) and does the estimation math itself; the client LLM writes the narrative, this server just supplies grounding and numbers.

This package is the MCP companion to QAI Consultant, an AI QA Architect web app / CLI. If you're looking for the full app (Risk Register / Effort Estimation / Test Strategy / Test Plan generation with a browser UI), see the main project instead — this package is just the MCP server piece of it.

Install

Requires uv (for uvx).

uvx qai-consultant-mcp

Claude Code:

claude mcp add qai-consultant -- uvx qai-consultant-mcp

Claude Desktop (claude_desktop_config.json):

{
  "mcpServers": {
    "qai-consultant": {
      "command": "uvx",
      "args": ["qai-consultant-mcp"]
    }
  }
}

First run downloads the embedding model (sentence-transformers/all-MiniLM-L6-v2, served via fastembed's ONNX Runtime) and builds a local index — a few seconds to a minute the first time, then it's cached.

Optional agent skill (Claude Code): copy the skills/qai-consultant/ folder into ~/.claude/skills/ (or your project's .claude/skills/). It tells the agent which tool fits a request and how to read the results, for example that an empty flaky-test list after only 1-2 runs means "not enough data", not "no flaky tests". Other agents that read SKILL.md files can use the same folder.

Tools

Try It — Example Prompts

Type these directly in Claude Code or Claude Desktop once the server is attached:

• "Using qai-consultant, what does the knowledge base say about risk-based testing?" → retrieve_qa_knowledge
• "List every document in the qai-consultant knowledge base, grouped by category." → list_kb_sources
• "Using qai-consultant's estimate_qa_effort, estimate the effort for a B2B SaaS project called 'Test Migration', React + Node.js + PostgreSQL, 2 QA / 5 dev engineers, 3-month timeline, Agile/Scrum, no notable known risks, minimal existing test automation, no special compliance requirements." → estimate_qa_effort
• "Here's a Test Plan [paste it] — use qai-consultant to review it with review_qa_document and give me the score and findings." → review_qa_document
• "Here's a JUnit XML report from my last 3 CI runs [paste them] — use qai-consultant's analyze_test_results to find flaky tests." → analyze_test_results
• "Here's how our team tests today [describe it] — use qai-consultant's assess_qa_maturity to estimate our TMMi level and the gaps to close." → assess_qa_maturity

The first one (retrieve_qa_knowledge) is the fastest way to confirm the server actually attached.

Prompts

• qa_project_interview — the project-intake interview (11 questions covering scope, tech stack, team, timeline, risks, compliance)
• risk_register_structure — Risk Register document structure + grounding instructions
• test_strategy_structure — Test Strategy document structure + grounding instructions
• test_plan_structure — IEEE 829-aligned Test Plan structure + grounding instructions

Each *_structure prompt instructs the client to ground its generation in retrieve_qa_knowledge results with [Source N] citations, and to label the output as AI-generated.

Privacy

Anonymous usage telemetry is on by default (since v3.7.0). Turn it off with QAI_TELEMETRY=0 or DO_NOT_TRACK=1; the server prints a one-line notice to stderr at startup whenever it's on. Only the tool name, a success flag, duration, retrieval k and category (only when it is one of the five known categories), package/Python version and OS family are sent, under a random ID that is new for every server run and never written to disk — never your query text, project details, or knowledge-base content.

Source

github.com/gvasile29/qai-consultant — Apache 2.0 licensed.