DiffCI MCP server

DiffCI

Find test-selection opportunities in your CI before changing what it runs. DiffCI analyzes a commit's changes and dependency graph, then reports which test files it would select, why it falls back to a full run, and whether it can propose a test command. check also runs paired full and selected commands when it can infer them. The observe command and Action remain observation-only.

Try DiffCI

From a Git repository checkout, with Node.js 22.5+ and Git installed, run:

npx @diffci.com/diffci@latest check

check explains affected tests and, when it can infer safe commands, runs both the full and selected test commands to measure the difference. It sends nothing to DiffCI. Test commands can create files in the checkout. For analysis without test execution, use npx @diffci.com/diffci@latest observe --no-send.

To add instructions for coding agents, run:

npx @diffci.com/diffci@latest init

On Windows PowerShell, quote the package name:

npx '@diffci.com/diffci@latest' check

The copyable adoption kit includes an AGENTS.md instruction and maintainer PR text. AI-readable documentation is on Context7 CLI and Context7 Core. Use the GitHub Marketplace Action for a separate, non-blocking observation job. Required project CI remains authoritative.

One paired run is preliminary evidence; repeat comparisons and account for cache effects before claiming CI savings. On a full-validation fallback, check runs the full command once and reports 0% reduction.

Upgrade from 0.1.3: tests excluded by a source-only tsconfig.json could be discovered without their dependency edges, producing an incomplete selection. This is fixed in 0.1.4. Revalidate affected observations before using them as opportunity evidence; see the historical validation and release qualification.

The local default compares HEAD with its first parent; both commits must be available. For a specific comparison, add --base --head . DiffCI prints the selection, fallback reasons, and the path to a JSON report outside your checkout. REFUSED or ERROR is not a successful analysis; check the reported status even when the command exits successfully. See the support matrix for setup requirements and supported workloads. check runs inferred full and selected commands in the checkout and sends nothing by default. The commands may write generated files. Use observe --no-send for analysis without execution. See docs/ai-agents.md for Claude Code, Codex, Cursor, GitHub Copilot, and similar tools.

Measured example: a controlled Cal.com replay showed 44.2% net reduction in a job-equivalent install + pretest + test workload, including analysis overhead. This is one sandbox comparison, not Cal.com's production savings or a prediction for your repository. Read the timings and method.

Shortened. The full README is on GitHub.