Open source for the era of AI agents.
Vyuh Labs builds systems where agents act with context, bounded capabilities, deterministic checks, and audit trails. Our open-source work brings those primitives to developers, starting with dxkit.
dxkit: context to make the change, a gate to stop cleanly.
AI coding agents can move fast, but they need more than a prompt and a test suite. They need structural context while they work, and a deterministic stop condition before they declare done. dxkit gives agents both: a repo code graph for structural context and a deterministic stop gate that blocks net-new regressions before the loop exits.
dxkit
Deterministic stop gate and code graph for AI coding agents.
Agents can pass tests and still leave a repo worse. dxkit baselines existing findings, gives the agent structural code context, and blocks only net-new regressions introduced by the current change.
The verification layer should be inspectable.
Agentic software development is moving from suggestions to action. That changes the safety problem. When an agent can edit files, run tools, and decide when it is done, the checks around that loop need to be visible, reproducible, and testable. Vyuh Labs open-sources the pieces that should not be hidden behind a black box: deterministic gates, benchmark harnesses, structural context tools, and reproducible agent-loop experiments.
Deterministic where it matters
The model can write and repair code. The gate that decides whether the loop may stop should be deterministic, local, and repeatable.
Built for brownfield repos
Real codebases already have debt. Open-source tools should baseline the current state and block new regressions without pretending existing complexity does not exist.
Research-backed
We publish methodology, caveats, and reproductions so developers can inspect what the claims mean and where they stop.
What AI coding agents do when nobody is watching.
dxkit grew out of a measurement project: run autonomous coding loops, inspect the final tree, and ask whether the agent made the repo worse before declaring success.
Try it, break it, and tell us where agents need a stronger gate.
We are especially interested in feedback from developers using Claude Code, Codex, Cursor, Aider, custom coding loops, and CI-driven repair agents. What should dxkit support next: more scanners, better code graph context, CI integration, pre-push workflows, or richer specification checks?