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:

Tool	What it does	Key needed
`codna_triage`	Understand a repo and locate the code relevant to an issue. Deterministic, 0 LLM tokens.	none
`codna_secure`	Prove which SARIF scanner findings (CodeQL, Semgrep, Snyk, Trivy) are reachable. Read-only, 0 LLM tokens.	none
`codna_recall`	Recall code from local on-device memory — semantic + lexical search, fully offline.	none
`codna_fix`	Root-cause and plan a fix (read-only by default); with `open_pr=true` pushes a branch and opens a real PR.	provider key (+ `GITHUB_TOKEN` for `open_pr=true`)
`codna_report_bug`	File a bug, feature, or question to thyn-ai/feedback.	`GITHUB_TOKEN` (else returns a pre-filled URL)

3 of 5 tools need no key at all — codna_triage, codna_secure and codna_recall run with zero credentials and zero network calls. Full reference: docs.codna.ai/guides/mcp.

Code memory

Code memory and recall run on your machine with no login and no extra install. 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.

Release files for codna 0.2.83

For a detailed explanation of source distributions (sdists) and built distributions (wheels), please see the package formats documentation.

Source distribution (sdist)

Source distribution for codna 0.2.83
File	Size	Uploaded
codna-0.2.83.tar.gz	3.9 MB	Sep 23, 2026	Details

Built distributions (wheels)

Table of built distributions (wheels) for codna 0.2.83
File	Interpreter	ABI	Platform	Reset
codna-0.2.83-py3-none-manylinux_2_28_x86_64.whl 64.2 MB Sep 23, 2026	Python 3	none	Linux glibc 2.28+ x86-64	Details
codna-0.2.83-py3-none-macosx_14_0_arm64.whl 29.7 MB Sep 23, 2026	Python 3	none	macOS 14.0+ ARM64	Details

Total release size: 97.8 MB

Release files / codna-0.2.83.tar.gz

Download URL	codna-0.2.83.tar.gz
Size	3.9 MB
Tags	Source
SHA-256 checksum How to use checksums	`eee031a2273b1bb5315bdd1d4ecd5178f0dfbce316724aeba384e9dbdafab0ff`
BLAKE2b-256 checksum How to use checksums	`d54b583db1ddcf80a55e31cb360a6b231b2c71233f817e77a4825378a36a9318`
Upload date	Sep 23, 2026
Uploaded using Trusted Publishing? What is trusted publishing?	No
Uploaded via	`twine/7.0.0 CPython/3.14.7`

Release files / codna-0.2.83-py3-none-manylinux_2_28_x86_64.whl

Download URL	codna-0.2.83-py3-none-manylinux_2_28_x86_64.whl
Size	64.2 MB
Tags	Linux glibc 2.28+ x86-64 Python 3
SHA-256 checksum How to use checksums	`debe3fb996d47bd4c04c854c84bf4f68a17daa2dfc236c811fb7317ae319d5ab`
BLAKE2b-256 checksum How to use checksums	`2a9bd9b83948018c495cbbe3aef12fcc64fc3aa8d0785583d83b6b14e7c6e35b`
Upload date	Sep 23, 2026
Uploaded using Trusted Publishing? What is trusted publishing?	No
Uploaded via	`twine/7.0.0 CPython/3.14.7`

Release files / codna-0.2.83-py3-none-macosx_14_0_arm64.whl

Download URL	codna-0.2.83-py3-none-macosx_14_0_arm64.whl
Size	29.7 MB
Tags	Python 3 macOS 14.0+ ARM64
SHA-256 checksum How to use checksums	`f45defce4eadb1f4312f1251a5b24e3051febf4f4151f662f08b6aef98d3c3ed`
BLAKE2b-256 checksum How to use checksums	`7a44d35277f3ace0a0caa55df7217e76beec13fa45c83abda19dd77d6381ddaa`
Upload date	Sep 23, 2026
Uploaded using Trusted Publishing? What is trusted publishing?	No
Uploaded via	`twine/7.0.0 CPython/3.14.7`