无需模型 token 的代码库理解。
Codna 将符号、导入、调用路径、测试和依赖解析为智能体可查询的实时依赖图。
Symbols mappedevery symbol
Languages250+ supported
LLM 消耗0 tokens
产品
智能体接收的是证据，而非整个代码库。
Codna 不向上下文窗口堆积文件，而是创建紧凑的证据包：可疑文件、调用链、失败测试和风险地图。
bundle: failing_test: checkout.spec.ts suspect_files: 4 call_paths: 7 estimated_context: ~600 tokens
核心能力
从问题到经验证的 Pull Request。
1
分类
在毫秒内理解任意本地路径或 Git URL，并定位可能需要变更的位置。
2
修复
生成携带根因分析、置信度评分和回归风险估算的补丁。
3
审查
结合影响范围、涉及测试和 API 影响审查生成的变更。
4
PR
使用 GitHub App 开启附带证据的经验证修复 Pull Request。
Distribution
Three ways to run Codna.
CLIRun codna fix in any repo, CI job, or container.
MCP serverGive Cursor and Claude codebase understanding as a local tool.
GitHub AppTriage issues and open verified fix pull requests.
Frequently asked
A deterministic engine builds a dependency and blast-radius graph in about 60ms, using zero LLM tokens. That graph produces a focused ~600-token evidence bundle — 162x less context than reading the repository — so the AI agent works only on what matters.
Every fix is verified by your own test suite before it ships. Nothing merges until your tests pass.
Codna supports 250+ languages, and has mapped 130 repositories in 9.2 seconds for zero tokens. If your project has tests, Codna can work with it.
In head-to-head testing across 87 tasks, Codna used 5× fewer tokens than Cursor and ran 1.7× faster, with every fix verified by the project's own tests (87/87). Both agents were measured on the same tasks.
Codna ships as a CLI, an MCP server that works inside Cursor and Claude, and a native GitHub App that opens verified fix pull requests directly in your repo.
No. You can self-host Codna, bring your own API key, and egress is fail-closed. Your code is never used for training.
