Integrations

Codna works where agents work.

One engine, everywhere your team already works — editor, terminal, CI, and GitHub. Every fix is verified by your own tests before it lands.

DocsContact

Integrations

Cursor

Add Codna as an MCP server. Cursor's agent gets your repo mapped in ~60ms for zero tokens — then fixes from a ~600-token evidence bundle.

Claude

Expose Codna's tools to Claude over MCP. The agent works from focused evidence — measured 162× less context than reading the repo.

GitHub

Native app that triages issues and opens fix PRs — and only opens one after your test suite passes.

CI

Run Codna in any CI job. It maps the failure, generates a patch, and verifies it against your tests — in the same run.

Model providers

Your agent. Your provider. Your key.

Codna maps your repo deterministically for zero tokens. The agent then runs on the model you choose — bring your own key, or use the managed LLM.

CLIRun codna fix in any repo, CI job, or container.
MCP serverGive Cursor and Claude codebase understanding as a local tool.
GitHub AppTriage issues and open verified fix pull requests.

Bring-your-own-key works today with an Anthropic (Claude) API key. Support for more model providers is on the way.

Frequently asked

Codna ships as an MCP server. Add it once and it becomes a native code agent inside Cursor and Claude — the same deterministic graph engine and AI fix pipeline, available right where you write code.

Yes. The GitHub App detects a bug, runs the engine, verifies the fix against your tests, and opens a pull request. Every PR it creates has passed your test suite before you see it.

Yes. The CLI integrates into any CI pipeline. It maps the affected graph, generates a fix, and verifies it — all within the same job, with no external state required.

No. Codna is bring-your-own-key. You supply the model API key, and the fix runs against your chosen model. Codna does not call a hosted LLM on your behalf or retain your code.

You run the engine on your own infrastructure. Egress is fail-closed, meaning no data leaves unless you explicitly configure it. Codna does not train on your code.

Very lightweight. The engine maps a repository in roughly 60ms using zero LLM tokens. The AI agent then works from a ~600-token evidence bundle, measured at 162x less context than reading the full repo.