Codna

Understand. Fix. Evolve.

Agents read your code. Codna understands it.

Codna maps a repository before it spends a token. It then reviews pull requests, fixes bugs and proves which scanner findings are reachable. The same codna command runs on your machine, in the GitHub Action and behind the GitHub App.

Runs on your machine or your cloud. Your code never leaves without your key.

Install

pip install codna          # or: pipx install codna · uv tool install codna
codna --version

Python 3.12–3.13. Wheels for macOS on Apple silicon and Linux x86_64. git on your PATH. Nothing else to install: no Node, Bun, Docker or server. The agent runtime ships inside the wheel.

Local commands need no Codna key. fix and review use your own model provider key, stored in the OS keychain and never printed:

codna key set anthropic    # also: openai, gemini, google, groq, mistral, openrouter, xai, cursor

Commands

codna triage . --issue "checkout total is wrong"     # suspect files. Deterministic. 0 LLM tokens.
codna review . --pr 123 --post                        # findings with a verdict on the pull request
codna fix . --tests --apply --max-iterations 3        # patch, re-run your tests, re-fix until green
codna fix <git url> --ref <sha> --issue "…" --open-pr # push a branch and open a pull request
codna secure . --from-sarif results.sarif             # which scanner findings are reachable. 0 LLM tokens.

codna fix prints the root cause, the impacted symbols, the blast radius, its confidence and a regression risk. --open-pr needs a git URL and a GitHub write token. codna review posts one inline comment per finding with severity, category and a suggestion block, and an Approve when the diff is clean at medium and high. codna secure reads SARIF 2.1.0 from CodeQL, Semgrep, Snyk, Trivy or any other scanner.

Other commands: init, status, doctor, login, key, impact, memory export, report. Full reference: docs.codna.ai/reference/cli.

MCP server

Run Codna as a Model Context Protocol server over stdio — the same engine the CLI uses, inside Cursor, Claude Desktop, or your own agent:

pipx install "codna[mcp]"       # or: pip install "codna[mcp]"
codna mcp                       # serve over stdio
codna mcp install --client cursor     # optional: write the client config for you (or --client claude)

Five tools, each returning JSON; a failure comes back as codna_<tool> error: … text and never crashes the server:

Introspection (initialize/tools/list) needs no credentials. Executing any of the five tools requires the one-time free community login (codna login — device authorization, free community license) — fully offline thereafter. codna_fix additionally needs a provider key (BYOK, e.g. ANTHROPIC_API_KEY); codna_report_bug needs GITHUB_TOKEN or it returns a pre-filled issue URL; codna_recall additionally uses the on-device memory runtime that the same codna login installs. Full reference: docs.codna.ai/guides/mcp.

Code memory

Code memory and recall run on your machine after a one-time free codna login (device authorization), which also installs the signed on-device runtime. From then on, recall runs fully offline: no key, no network calls. The optional codna[memory] extra adds the on-device semantic reranker; without it, recall ranks lexically.

What leaves your machine

• Repository mapping, triage, recall and impact run offline. No model is involved.
• fix and review send one issue-specific evidence bundle to the provider you chose, under your key. Not the repository.
• Secret redaction is always on and cannot be turned off.
• privacy.egress: fail-closed in codna.yaml refuses to run tests without network denial and skips registry lookups during review.
• Source distributions exclude runtime binaries, keys, .env files and logs.

Links

• Homepage: https://codna.ai
• Documentation: https://docs.codna.ai
• Source (public mirror): https://github.com/thyn-ai/codna-cli
• Security: https://codna.ai/security