1
Map for zero tokens
Codna builds a dependency and blast-radius graph of the repository from its import patterns. No model call, zero tokens.
Comparison
Codna vs Claude Code
Claude Code is an agent. Codna is the understanding it works from: a deterministic map for zero tokens, an evidence bundle scoped to the issue, and a fix that explains itself. It plugs in over MCP.
The problem
A strong agent still has to find the bug.
Reading the repository through the model is orientation cost. Codna removes it: the graph is built deterministically, recall runs on-device, and the agent starts with the evidence in hand.
How Codna fixes it
How Codna works
2
Hand over evidence, not the repo
The agent receives an evidence bundle scoped to the issue: the suspect files, the call paths, the failing test. Codna prints the raw-to-bundle token size on every run.
3
Report, gate, open the PR
Every fix reports root cause, confidence, blast radius and regression risk, and passes a risk gate before it is applied or a pull request opens. With --open-pr, or through the GitHub App, the pull request states the issue, the root cause, the symbols touched and a confidence score, and asks for review before merging. Codna never merges.
pip install "codna[mcp]" codna mcp install --client claude
What you get
What you get
A zero-token repo map
Codna builds a dependency and blast-radius graph of the repository from its import patterns. No model call, zero tokens.
Five tools in Claude
codna_triage, codna_fix, codna_secure, codna_recall and codna_report_bug, installed with one command and without writing credentials.
A fix that explains itself
Every fix reports root cause, confidence, blast radius and regression risk, and passes a risk gate before it is applied or a pull request opens.
The proof
Fewer tokens. Faster. Verified.
Frequently asked
The only measured head-to-head Codna publishes is against Cursor: 87 matched bug-fix cases on identical checkouts, 5× fewer tokens, 1.7× faster, 87 of 87 verified, about $0.02 per fix. Everything else on this page describes Codna's own behaviour, not Claude Code's internals.
No. Codna runs alongside it. Install the mcp extra and run codna mcp install --client claude; Claude gets the five Codna tools as a local server.
Codna separates understanding from fixing. Codna builds a dependency and blast-radius graph of the repository from its import patterns. No model call, zero tokens. The agent receives an evidence bundle scoped to the issue: the suspect files, the call paths, the failing test. Codna prints the raw-to-bundle token size on every run. The agent, Claude or another, works from that bundle.
Yes. Codna is a CLI, an MCP server, a GitHub Action and a GitHub App. Add it where it helps and keep the rest of your workflow.
Bring a key from Anthropic, OpenAI, Google Gemini, Groq, Mistral, OpenRouter or xAI, stored in your OS keychain with codna key set, or sign in and use the managed allowance. The repository map is built from import patterns, so it is not tied to one language. On-device recall covers Python, JavaScript, TypeScript, Go, Rust, Java, C, C++, C#, PHP and Ruby. Your own test command verifies the fix.
Understanding runs on your machine and spends no tokens. Only the evidence bundle or the diff reaches your model provider, with your key from the OS keychain. Set privacy.egress to fail-closed in codna.yaml and Codna runs your tests only under kernel-level network denial. Secret redaction is always on. On 87 matched bug-fix cases against Cursor, Codna averaged 16,159 total tokens and about $0.02 of model spend per verified fix, 5× fewer tokens and 1.7× faster. With your own key you pay your provider directly.
Related
