A readiness readout, not a vibe
Point it at a repo.
Get the readout.
Latch reads a repo the way an AI agent must — top to bottom, with no one to ask — and reports a score per signal, plus the precise fixes that raise it.
Six categories · 600-point scale · receipts included.
// the six signals we read
The six signals we read
Each signal scored 0–100, with a one-line readout that doesn't round up.
Machine Orientation
Whether an AI agent can grasp what the repo is and how to work in it from concise, machine-readable context — a README and an AGENTS.md/CLAUDE.md stating build/test/run commands, entry points, and conventions — not padded or auto-generated volume.
Reproducible Setup
How deterministically an agent can install, build, and run the project — pinned versions, lockfiles, a container or dev-container, and explicit commands — rather than guessing at the environment, since nothing downstream works until it runs.
Code Navigability
Whether an agent can localize the exact code to change — clear module boundaries, consistent, descriptive naming, typed signatures, and files that fit a context window — instead of editing the wrong place.
Verification Surface
Whether an agent can check its own work against ground truth — runnable, trustworthy tests (that actually pass and aren't flaky), type-checking, linters, and CI it can read — closing the feedback loop instead of editing blind.
Agent Safety
How safe it is to let an autonomous agent act — secret hygiene (a .env.example, never a committed .env), flagged destructive or irreversible operations, prompt-injection surfaces in agent-read content, and clear 'don't touch' boundaries.
Context Economy
Whether the signal an agent needs fits its effective context window — signal-to-noise, repo size, and generated or vendored bulk marked as such rather than drowning the useful parts.