1
Map for zero tokens
Codna builds a dependency and blast-radius graph deterministically — no LLM, ~60ms per repo.
Comparison
Codna vs Codex CLI
Both can fix a bug. The difference is what they read first. The Codex CLI explores your repo inside the model; Codna maps it deterministically for zero tokens, then fixes from a tiny evidence bundle.
The problem
Where the Codex CLI spends its budget
The Codex CLI is a capable agent, but to understand a repository it reads file after file inside the model — tens of thousands of tokens before it even proposes a fix. That cost scales with repo size and bug count, and it has to repeat the exploration on the next task.
How Codna fixes it
How Codna fixes the same bug differently
Whole repo100,000+
Evidence bundle~600
2
Hand over evidence, not the repo
The agent receives a ~600-token bundle: suspect files, call paths, the failing test.
3
Verify before it ships
The patch must pass your test suite. Across 8 head-to-head scenarios, every Codna fix passed.
codna fix . --issue "the checkout test is failing"
What you get
Why teams pick Codna
Understanding for zero tokens
Codna maps the repo into a dependency and blast-radius graph deterministically — no embeddings, no RAG, about 59ms per repo (p50) for 0 LLM tokens. The Codex CLI explores the same repo inside the model, spending tens of thousands of tokens before it proposes a fix.
Up to 18x fewer tokens, measured
Across 8 head-to-head bug-fix scenarios on identical checkouts, Codna used 12.9x fewer tokens and ran 5.2x faster than the Codex CLI in aggregate — up to 18x fewer tokens on the best scenario.
Every fix gated by a passing test
A patch only lands after it passes your own suite, so a fix is proven rather than plausible. In the head-to-head, every Codna fix passed its oracle test — 8 of 8.
The proof
Fewer tokens. Faster. Verified.
Frequently asked
Measured. We ran 8 bug-fix scenarios on identical isolated checkouts on 2026-06-15, each verified by a passing oracle test for every agent. In aggregate Codna used 12.9x fewer tokens and ran 5.2x faster than the Codex CLI, with the best single scenario at up to 18x fewer tokens.
It can do either, and it most often complements it. Codna is the deterministic understanding layer plus a test-gated fix agent; you can feed its evidence bundles to the Codex CLI through the MCP server and keep your existing workflow. Many teams use Codna for repo-wide understanding and verified fixes, and the Codex CLI for open-ended coding.
The Codex CLI is an open-source terminal agent that explores the repository inside the model, reading file after file to orient itself. Codna maps the repo into a dependency and blast-radius graph deterministically first — no LLM, no embeddings — then hands the agent a roughly 600-token evidence bundle instead of the whole repo. That is why the same fix costs far fewer tokens and less wall-clock time.
Yes. Codna is model-agnostic and bring-your-own-key, and it ships as a CLI, an MCP server, and a GitHub App. Pair its evidence bundles with whatever model and agent you already run, including the Codex CLI via MCP.
Codna builds its graph from the source itself, so it spans many ecosystems — 130 repos across 110 languages mapped in 9.2 seconds at 100% ecosystem accuracy in our run. The verified bug-fix scenarios covered Python and TypeScript, including never-before-seen repositories.
Only if you choose to. Codna is bring-your-own-key and can self-host with fail-closed egress and no training on your code. Because the agent fixes from a roughly 600-token bundle rather than reading the whole repo, a verified fix costs about $0.04 measured at public model rates — versus around $1 for a typical agentic edit that reads everything.
Related
