Docs

Install Codna in one command.

Get Codna running in minutes — CLI, MCP server, or GitHub App, your choice.

InstallGitHub App

Install

Install the CLI locally. Codna can run on your machine, your cloud, or self-hosted infrastructure.

pip install codna
codna --version

Quickstart

First, understand a repo. Then ask Codna to fix a specific issue.

codna triage .
codna fix . --issue "the checkout test is failing"

CLI

Core commands are intentionally small and scriptable.

codna triage . --json
codna fix . --issue "..." --tests
codna fix https://github.com/org/repo --issue "..."

MCP server

Add Codna to Cursor or Claude in one line. Your agent can ask Codna for symbol maps, dependencies, test impact, and focused evidence bundles.

codna mcp start --repo .

Native GitHub App

Install the GitHub App to let Codna triage issues, comment with root-cause evidence, and open fix pull requests.

# App flow
Install codna → select repos → enable fix PRs

Configuration

Bring your own model key, or use Codna's managed model — it's the same Codna either way.

model:
  provider: openai
  key: env:CODNA_MODEL_KEY
privacy:
  egress: fail-closed
  redact_secrets: true

FAQ

Does Codna train on my code? No. Codna is designed for no training on customer code.

Can I self-host? Yes. Self-hosting is available forever, with supported on-prem and air-gapped options for Enterprise.

Frequently asked

Install with a single command from your terminal. The codna cli runs locally, so you bring your own API key and your code never leaves your machine unless you choose otherwise.

The MCP server for coding integrates natively with Cursor and Claude. Add it to your editor's MCP config and Codna's graph engine becomes available to your AI assistant without any extra steps.

The GitHub App monitors your repository and opens pull requests with verified fixes automatically. Every fix is validated by your own test suite before the PR is created.

No. Codna is self-hostable, and egress is fail-closed by default. You supply your own API key, and the system is designed so your code is never used for model training.

The deterministic graph engine maps the relevant code first, producing an evidence bundle measured at around 600 tokens. Head-to-head against Cursor, Codna used 5× fewer tokens and ran 1.7× faster, every fix test-verified (87/87).

Codna has mapped 130 repositories, completing the full mapping in 9.2 seconds for zero LLM tokens. If your project compiles or resolves dependencies, Codna can graph it.