What is Forgecroft?

Forgecroft is a verification layer for autonomous engineering. It structures your real systems — code, infrastructure, configuration — into a graph of Systems, Components, and Controls, then evaluates each change against that graph before it merges. Green is not safe: a passing test suite says the code runs, not that the change respects your organization’s contracts. Forgecroft answers the second question, with calibrated honesty about how much is actually known.

The fastest path to a verdict is the CI gate and the MCP self-correction loop — both call forgecroft_evaluate on your diff and return a deterministic decision (go, conditional_go, no_go, or insufficient_context) with no model on the read path.

Gate a change in CI

Drop the Forgecroft gate into your repository so every pull request is evaluated before merge. The gate runs inside your own CI runner and sends only the changed diff to the Forgecroft API over TLS — your source tree never leaves your environment.

See CI Gate for the GitHub Actions workflow, fail-open semantics, and offline mode.

Self-correct from a coding agent

Wire the Forgecroft MCP server into Claude Code, Codex, or Cursor and have the agent call forgecroft_evaluate on its diff mid-loop, before pushing. Each finding carries a fix_hint the agent can act on immediately.

See Coding Agents for MCP configuration and the self-correction workflow.

What you need

• A repository with code or configuration you want gated
• A Forgecroft API key with evaluate:read scope (or run --offline against the embedded best-practice packs with no key)

How it works

forgecroft_evaluate matches your changed file paths to the covering Components in your org’s graph, applies your Controls (deterministic compiled checks, not an LLM judge), and returns the verdict with blast_radius, confidence, and coverage_notes. Seed your graph for richer blast-radius — see Core Concepts for the model.