# Codna > Codna is code intelligence for AI agents: a deterministic engine maps an entire repository for zero model tokens, then hands an agent an evidence bundle scoped to the issue and prints the raw-to-bundle size on every run. Every fix reports root cause, confidence, blast radius and regression risk and passes a risk gate before it is applied. Codna also reviews pull requests with a clear verdict and proves which scanner findings are reachable. Measured against Cursor on 87 matched bug-fix cases, Codna used 5× fewer tokens and ran 1.7× faster, at about $0.02 of model spend per verified fix. Operated by Seville AI Technologies, Inc. Runs on your machine with your own key (Anthropic, OpenAI, Google Gemini, Groq, Mistral, OpenRouter or xAI) or on a managed allowance. Secret redaction is always on; fail-closed egress is an opt-in codna.yaml setting. Your code is not used to train models unless you opt in. The site is available in 13 languages. ## Pages - [Product](https://codna.ai/product): How Codna maps a repo deterministically for zero tokens, then fixes, reviews and secures with every result explained. - [How it works](https://codna.ai/how-it-works): Understand → fix → gate: the deterministic map, the evidence bundle, and the risk gate before anything is applied. - [Benchmarks](https://codna.ai/benchmarks): Measured against Cursor on 87 matched bug-fix cases: 5× fewer tokens, 1.7× faster, 87 of 87 verified, about $0.02 per fix. - [Comparisons](https://codna.ai/compare): Codna vs Cursor, Codex, Gemini, Claude Code and more — measured where a measurement exists. - [Use cases](https://codna.ai/use-cases): Fix flaky tests, turn Sentry issues into PRs, repair CI failures, prove scanner findings reachable, and more. - [Integrations](https://codna.ai/integrations): A CLI, an MCP server, a GitHub Action and a GitHub App — the same codna commands everywhere. - [GitHub App](https://codna.ai/github-app): Reviews every pull request with a clear verdict; opens fix PRs from an issue label, a comment or a red check. Never merges. - [Languages](https://codna.ai/languages): The repository map is built from import patterns in any language, for zero model tokens. - [Security](https://codna.ai/security): Runs on your machine with your own key. Only the evidence bundle reaches your provider. Redaction always on; fail-closed egress when you set it. - [Enterprise](https://codna.ai/enterprise): Review on every PR, fix PRs you approve, least-privilege App tokens, and a server licence for your own production deployment. - [Pricing](https://codna.ai/pricing): Free on your machine with your own key; managed plans when Codna should carry the model bill. - [Docs](https://codna.ai/docs): Install with pip, map a repo, fix, review and secure from the CLI, MCP, the Action or the GitHub App. - [Resources](https://codna.ai/resources): Guides, the measured benchmark, and head-to-head comparisons. - [Full documentation](https://docs.codna.ai): The complete Codna reference — quickstart, CLI, MCP server, GitHub Action, configuration, and concepts. ## Comparisons - [Codna vs Codex CLI](https://codna.ai/compare/codex): Codna maps the repository for zero tokens and hands the agent an evidence bundle. The Codex CLI is an agent you drive from the terminal. See how the approaches differ. - [Codna vs Gemini CLI](https://codna.ai/compare/gemini): Codna maps the repository deterministically for zero tokens, then fixes from an evidence bundle. The Gemini CLI is an agent driven from the terminal. Compare the approaches. - [Codna vs Cursor](https://codna.ai/compare/cursor): Measured on 87 matched bug-fix cases: Codna used 5× fewer tokens and ran 1.7× faster than Cursor, with 87 of 87 fixes verified. Codna also runs inside Cursor as an MCP server. - [Codna vs Claude Code](https://codna.ai/compare/claude-code): Codna gives Claude Code a deterministic repository map and an evidence bundle over MCP. Five tools, zero tokens to understand, every fix explained. - [Codna vs Devin](https://codna.ai/compare/devin): Devin is an autonomous software agent. Codna maps the repository deterministically for zero tokens, fixes from evidence, and opens a pull request you review. Compare the approaches. - [Codna vs Windsurf](https://codna.ai/compare/windsurf): Windsurf is an agentic editor. Codna is a deterministic understanding layer: zero tokens to map the repo, an evidence bundle for the agent, and a fix with root cause and risk attached. - [Codna vs Augment Code](https://codna.ai/compare/augment): Augment Code is an AI coding platform. Codna builds a deterministic dependency graph for zero tokens, hands the agent an evidence bundle, and reports root cause and risk on every fix. - [Codna vs Sourcegraph Amp / Cody](https://codna.ai/compare/amp): Sourcegraph Amp and Cody are coding agents with code search behind them. Codna maps the repository deterministically for zero tokens and reports root cause and risk on every fix. - [Codna vs Greptile](https://codna.ai/compare/greptile): Greptile reviews pull requests. Codna reviews them with severity, confidence and a clear verdict, then opens the fix when you reply @codna fix. Compare the approaches. - [Codna vs CodeRabbit](https://codna.ai/compare/coderabbit): CodeRabbit reviews pull requests. Codna reviews them with severity, confidence and a clear verdict, then opens the fix when a collaborator replies @codna fix. - [Codna vs Aider](https://codna.ai/compare/aider): Aider is a terminal pair programmer. Codna maps the repository deterministically for zero tokens, fixes from evidence, and reports root cause, confidence and regression risk. ## Use cases - [Fix flaky tests with Codna](https://codna.ai/use-cases/fix-flaky-tests): Codna maps the suite for zero tokens, finds the shared state or ordering a flaky test depends on, and re-runs your tests until they pass. Root cause and risk on every fix. - [From Sentry alert to fix PR with Codna](https://codna.ai/use-cases/sentry-to-pr): Paste the Sentry title and stack trace into codna fix, or label the issue codna-fix. Codna maps the repo for zero tokens and opens a pull request with root cause and confidence. - [Fix CI failures with Codna](https://codna.ai/use-cases/ci-failures): A red check suite on a pull request triggers Codna. It triages the failure first, so infrastructure failures cost nothing, then opens a fix on the pull request branch. - [Fix monorepo bugs with Codna](https://codna.ai/use-cases/monorepo): Codna maps the whole workspace from its imports for zero tokens, scopes the fix to the package and its consumers, and lists the tests a diff can affect. - [Fix bugs in legacy code with Codna](https://codna.ai/use-cases/legacy-code): Codna maps legacy code from its imports for zero tokens, so you see what a change reaches before you make it. Every fix reports root cause, confidence and regression risk. - [Fix breakage from a dependency upgrade with Codna](https://codna.ai/use-cases/dependency-upgrade): A dependency upgrade breaks the build. Codna maps every call site from the import graph for zero tokens, fixes them with the consumers in view, and verifies with your tests. - [Prove and fix scanner findings with Codna](https://codna.ai/use-cases/security-fix): codna secure reads SARIF from CodeQL, Semgrep, Snyk or Trivy and proves which findings are reachable, for zero model tokens. Verified findings open as draft pull requests, one per PR. - [Fix a regression](https://codna.ai/use-cases/regression): Something that passed last release fails now. Codna maps the repository for zero tokens, fixes the cause rather than the symptom, and reports root cause, confidence and regression risk. - [Understand a new codebase with Codna](https://codna.ai/use-cases/understand-codebase): codna triage maps a repository for zero tokens and names the code an issue touches. codna_recall recalls code from an on-device index. Read-only, nothing changes. - [Fix Java bugs with Codna](https://codna.ai/use-cases/java): Codna maps a Java project from its imports for zero tokens, fixes from an evidence bundle, and runs your Maven or Gradle tests until they pass.