Developer tooling for AI-assisted software engineering.

We build tools that make AI coding agents safe to rely on — and we publish the evidence for what they do and don't do.

Declare each module's interface and dependencies in plain YAML. anma sync compiles that into the CLAUDE.md, hooks, and CI checks that keep an agent inside your architecture. Author once, enforce everywhere.

In a 20-trial benchmark, a cheaper model (Claude Haiku 4.5) violated a declared module boundary in 13 of 19 runs of a plain repo. With ANMA: 0 of 20 (Fisher's exact p < 0.0001). A frontier model respected the boundary on its own — so ANMA is insurance for running cheaper agents, plus a CI/governance guarantee. We publish the frontier null result alongside the win.

pip install anma[tach]
anma init && anma sync && anma check

→ Repository · Benchmark study · anmalabs.dev

Verified over claimed. Every performance number we publish is reproducible from a committed harness and raw data — including the cases where our tools make no difference. If we can't measure it, we don't assert it.

_{© 2026 ANMA Labs LLC · Apache-2.0}