Robot Framework MCP Server

A local Python MCP server for analysing Robot Framework output.xml results and exposing structured failure data to Claude Desktop and Claude Code.

What it does

• Parses Robot Framework output.xml files and persists results to a local SQLite database

• Scores every failing test across multiple dimensions (priority, severity, defect ID, quarantine, infra type, group size)

• Groups failures by shared root cause and ranks them into a triage queue

• Escalates API response-code failures (Wrong response code received) automatically — 5xx forced to score 92, other 4xx to 87, 403 to 80 — regardless of test tags

• Extracts the actual error from the response body JSON so you see the real server message, not just the status code

• Generates a prioritised failure report — interactive HTML by default (score-distribution doughnut chart, collapsible API failure groups, collapsible failure groups) or plain Markdown

• Prepares per-test debug context (keyword trace, innermost failure, product-file hints) linked to the robot-tests Claude skill

• Supports freeform SQL queries against the failure database for ad-hoc questions (results capped at 200 rows)

Architecture

Claude Desktop / Claude Code
        │
        ▼
examples/mcp_stdio_wrapper.py     ← FastMCP stdio transport (MCP protocol)
        │  http
        ▼
robot_mcp_server/http_server.py   ← FastAPI HTTP server (uvicorn, port 8000)
        │
        ▼
robot_mcp_server/
  failure_matrix.py   — scoring, grouping, triage queue, debug context
  db.py               — SQLite persistence (failure_analysis.db)
  parser.py           — output.xml loading, keyword trace extraction

Project structure

robot_mcp_server/       Python package
  http_server.py        FastAPI endpoints
  failure_matrix.py     scoring, triage, matrix rendering
  html_report.py        interactive HTML report renderer
  db.py                 SQLite persistence layer
  parser.py             output.xml parsing (RF 6.x compatible)
examples/
  mcp_stdio_wrapper.py  MCP stdio wrapper (connect this to Claude)
  sample_output.xml     minimal RF output.xml for testing
prompt_templates/
  review_robot_failures.md  Claude prompt template

Installation

python -m pip install -r requirements.txt
pip install mcp   # MCP Python SDK for Claude integration

Running

Start the FastAPI HTTP server (must be running before using any MCP tools):

cd /path/to/rf-test-analysis-mcp
python -m uvicorn robot_mcp_server.http_server:app --host 127.0.0.1 --port 8000

Check it's up: http://127.0.0.1:8000/docs

Connecting to Claude Desktop

Add to %APPDATA%\Claude\claude_desktop_config.json:

{
  "mcpServers": {
    "robot-framework-failure-review": {
      "command": "python",
      "args": ["C:\\path\\to\\rf-test-analysis-mcp\\examples\\mcp_stdio_wrapper.py"],
      "env": {
        "BASE_URL": "http://127.0.0.1:8000"
      }
    }
  }
}

Then restart Claude Desktop. The server appears under the + → Connectors menu.

Connecting to Claude Code

The .claude/settings.json in this repo configures the MCP server automatically for Claude Code when working in this directory.

Area category configuration (optional)

prepare_debug_context can route each failure to a named debug category and supply product-file hints for an external skill. This mapping is entirely optional — the server works without it, and all tests default to the GENERIC category.

To add your own mappings, create area_categories.json in the project root (or set RF_AREA_CATEGORIES to an absolute path):

[
  {
    "folder_fragment": "Checkout",
    "category": "CHECKOUT",
    "hints": ["checkout/checkout.md"]
  },
  {
    "folder_fragment": "Auth",
    "category": "AUTH",
    "hints": ["auth/auth.md"]
  }
]

Each folder_fragment is matched case-insensitively against the parent folder of each output.xml. An example with sample entries ships as area_categories.json.

Database initialisation

Before using get_triage_queue, generate_failure_matrix, or execute_query, you must ingest the results once:

ingest_results(results_dir="F:\path\to\results")

This parses all robot/output.xml files found in child folders, scores every failing test, and writes everything to failure_analysis.db inside results_dir. The ingest is idempotent — if nothing has changed (SHA-1 hash of all xml mtimes matches) it returns immediately. On a fresh run expect ~3 minutes for a large suite; subsequent calls complete in under 1 second.

You only need to re-run ingest after a new test run produces new output.xml files.

Available MCP tools

Failure scoring

Scores run 0–100. Higher = investigate first.

Typical workflow

1. Start FastAPI server (uvicorn)
2. Call ingest_results(results_dir=...) — one-time parse, ~3 min; instant if unchanged
3. Call generate_failure_matrix(output_path=...) → interactive HTML report written to disk
4. Call get_triage_queue → ranked groups, highest-impact first
5. For each group: call prepare_debug_context → keyword trace + routing hints
6. Use the robot-tests Claude skill (triage command) to drive the full session
7. Ad-hoc questions: call execute_query with a SELECT against scored_tests

Ad-hoc queries (execute_query)

Ask questions in plain English — Claude writes the SQL. Results are capped at 200 rows; the response includes truncated: true and returned when the cap is hit. Examples:

-- How many failures per area?
SELECT area, COUNT(*) AS failures FROM scored_tests GROUP BY area ORDER BY failures DESC

-- Which quarantined tests have the highest score?
SELECT name, area, base_score FROM scored_tests WHERE is_quarantined=1 ORDER BY base_score DESC

-- All API escalated failures grouped by endpoint
SELECT api_endpoint, COUNT(*) AS n, received_code
FROM scored_tests WHERE escalated=1
GROUP BY api_endpoint ORDER BY n DESC

-- Critical failures with no tracked defect
SELECT name, area, base_score FROM scored_tests
WHERE priority='critical' AND defect_ids='[]' AND is_quarantined=0
ORDER BY base_score DESC

Unit tests

pytest

Notes

• Compatible with Robot Framework 6.x (output.xml format changes from RF 4/5 handled)

• All analysis is read-only; no test results are modified

• The mcp_stdio_wrapper.py requires the FastAPI server to be running separately on port 8000

• The SQLite database (failure_analysis.db) lives inside the results directory alongside the xml folders