Blacksmith MCP

An MCP server that connects Claude to your Blacksmith CI data. Query workflow runs, analyze test failures, detect flaky tests, and monitor usage—all through natural conversation.

Why?

Debugging CI failures usually means clicking through dashboards, copying run IDs, and piecing together information across multiple pages. With this MCP, you can just ask:

"Why did the last CI run fail?"
"Which tests are flaky this week?"
"Compare test failures between main and my PR"
"What's using the most cache storage?"

Claude handles the API calls and gives you actionable insights.

Quick Start

Zero-config if you're logged into Blacksmith in Chrome:

server.ts

# Add to Claude Code
claude mcp add blacksmith -- npx blacksmith-mcp

# Set your org (run once)
export BLACKSMITH_ORG="your-org-name"

The MCP automatically extracts your session from Chrome cookies. No manual token copying needed.

Installation

Option 1: Claude Code CLI

Terminal

claude mcp add blacksmith -- npx blacksmith-mcp

Option 2: Project Configuration

Add to your .mcp.json:

config.json

{
  "mcpServers": {
    "blacksmith": {
      "type": "stdio",
      "command": "npx",
      "args": ["blacksmith-mcp"],
      "env": {
        "BLACKSMITH_ORG": "your-org-name"
      }
    }
  }
}

Option 3: Global Install

Terminal

npm install -g blacksmith-mcp

Configuration

Authentication

Automatic (recommended): Log into app.blacksmith.sh in Chrome. The MCP extracts your session cookie automatically.

Manual: Set BLACKSMITH_SESSION_COOKIE environment variable with your session cookie value.

Environment Variables

Variable	Required	Description
`BLACKSMITH_ORG`	Yes	Your Blacksmith organization name
`BLACKSMITH_SESSION_COOKIE`	No	Session cookie (auto-extracted from Chrome if not set)

Available Tools

Workflow Runs

Tool	Description
`list_runs`	List workflow runs with filters (status, branch, workflow, actor, PR)
`get_run`	Get run details including all jobs
`list_jobs`	List jobs for a workflow run
`get_job`	Get job details (steps, timing, runner info)
`get_job_logs`	Get raw log output for a job

Test Analytics

Tool	Description
`get_job_tests`	Get all test results for a job
`get_failed_tests`	Get failed tests with full error messages
`get_failures_by_pattern`	Group failures by error pattern (e.g., "Cannot read properties")
`compare_test_runs`	Compare failures between two runs (find regressions)
`get_flaky_tests`	Detect tests that fail intermittently
`get_slow_tests`	Find tests exceeding duration threshold
`get_test_history`	Track a specific test's pass/fail history
`get_trends`	Analyze trends: duration, failure rate, test count

Usage & Billing

Tool	Description
`get_current_usage`	Current core usage (active vs max)
`get_invoice_amount`	Current billing period amount
`get_usage_summary`	Billable minutes vs free tier
`get_cache_stats`	Cache size, entries by repository
`get_cache_entries`	Detailed cache entries for a repo

Organization

Tool	Description
`list_orgs`	List accessible organizations
`get_org_status`	Org status (personal, onboarded, region)
`search_logs`	Search logs across all jobs

Example Conversations

Debug a Failed Run

You: Why did the last CI run fail?

Claude: Uses list_runs(status="failure", limit=1) then get_failed_tests

The last run failed with 34 test failures. The main issues are:

Connection timeout (12 tests) - Tests in api/ are timing out connecting to the test database

Undefined property (8 tests) - user.settings.theme is undefined in the new dark mode code

Assertion failures (14 tests) - Snapshot mismatches in UI components

Find Flaky Tests

You: Which tests are flaky?

Claude: Uses get_flaky_tests(days=7, min_runs=5)

Found 3 flaky tests:

Test Flake Rate Pattern
should handle concurrent requests 40% (4/10 runs) Fails on CI, passes locally
renders loading state 20% (2/10 runs) Race condition with setTimeout
validates auth token 30% (3/10 runs) Intermittent network timeout

Test	Flake Rate	Pattern
`should handle concurrent requests`	40% (4/10 runs)	Fails on CI, passes locally
`renders loading state`	20% (2/10 runs)	Race condition with setTimeout
`validates auth token`	30% (3/10 runs)	Intermittent network timeout

Compare Branches

You: Did my PR introduce new test failures?

Claude: Uses compare_test_runs with PR run vs base branch

Your PR introduced 2 new failures:

UserProfile.test.tsx: "Cannot read property 'avatar' of undefined"

api/users.test.ts: Expected 200, got 404

5 previously failing tests now pass (nice!).

Development

bash

# Install dependencies
pnpm install

# Build
pnpm build

# Development mode (watch)
pnpm dev

# Test with MCP Inspector
npx @modelcontextprotocol/inspector node dist/index.js

Troubleshooting

Session Expired

If you see SESSION_EXPIRED, your Blacksmith session has expired. Simply log back into app.blacksmith.sh in Chrome and retry.

Cookie Extraction Failed

The automatic cookie extraction requires:

macOS with Chrome installed
Being logged into Blacksmith in Chrome
Chrome not running with a locked profile

If it fails, set BLACKSMITH_SESSION_COOKIE manually.

No Organization Set

Run list_orgs to see available organizations, then set BLACKSMITH_ORG to your org name.

API Notes

This MCP uses Blacksmith's internal web API, which is undocumented. The API was reverse-engineered from the Blacksmith web app and may change without notice.

License

MIT

Contributing

Contributions welcome! Please open an issue first to discuss proposed changes.