Today's coding agents brute-force your repo.
They read file after file inside the model — burning tokens, time, and context — just to figure out where the bug is. It is slow, expensive, and it forgets.
100Ktokens just to orient.
Codna maps your entire repo in milliseconds — for zero tokens — then sends an agent in with exactly the context it needs.
pip install codnaMeasured across 87 verified tasks — 5× fewer tokens and 1.7× faster than Cursor, 100% verified (vs Cline's 73.6%).
They read file after file inside the model — burning tokens, time, and context — just to figure out where the bug is. It is slow, expensive, and it forgets.
100Ktokens just to orient.
A deterministic engine maps every symbol, dependency, and blast radius — for zero tokens — then hands the agent a tight evidence bundle. Measured 162× less context than reading the repo.
~600tokens of focused context.
How it works
The agent does not wander through your repo. Codna gives it a map, evidence, and a safe path to ship.
Codna parses your repo into a dependency and blast-radius graph — no LLM, about 60ms per repo, zero tokens.
It pinpoints suspect files, packs a minimal evidence bundle, and runs the agent surgically with root cause and confidence.
Every fix is simulated for blast radius and risk before you apply it. Codna gets sharper the more you ship.
Benchmarks
Head-to-head on real public repos — single bugs, multiple bugs in one repo, and repos the agent had never seen. Codna used far fewer tokens and ran faster, and every fix passed the test.
| Codna | Codex CLI | Gemini CLI | |
|---|---|---|---|
| Tokens to fix · 8 scenarios | 465K | 6.0M | 4.5M |
| Time | 108s | 568s | 1,000s |
| Fixes verified | 8 / 8 | 8 / 8 | 8 / 8 |
Every scenario verified by a passing test. Full methodology and per-scenario data on the benchmarks page.
Workflow
One command, one GitHub app, or one MCP server. Codna sits where your engineering team already works.
codna fix . in any repo. Or codna triage . to understand it instantly.
codna triage . codna fix . --issue "checkout fails"
Add Codna to Cursor or Claude as an MCP server. Your agent gets a brain that actually knows the codebase.
Install Codna on your repos, triage issues, open fix pull requests, and keep engineers in the review loop.
Every fix comes with root cause, confidence, blast radius, and simulated regression risk.
Codna is built on a privacy-first engine: per-tenant key isolation, secret redaction, and audited egress. The guarantees are not a promise — they are enforced in the runtime.
Pricing
Understanding your repo is deterministic and costs almost nothing, so triage is unlimited on every plan. Agent fixes spend model tokens, so you pay when you scale.