Coding agents build fast. Cadora is the audit-grade conductor that proves what they built, checks nobody faked it, and tells you what it cost — on Claude Code, Codex, and more.
"Tests pass" means whatever the agent decided it means. A suite that runs zero tests still exits green.
Agents have been caught faking test runs and swapping in stub tools. Who checks the checker?
You ran it on Claude and Codex. Which was cheaper for this build? Every vendor's dashboard stops at its own fence.
Every agent vendor ships a way to run agents. None neutrally proves what its own agent did, sees across vendors, or hands you evidence you can forward.
Cadora re-runs your build and tests itself, reads exit codes and test counts, scans the toolchain for tampering, and captures every run. It never takes the agent's word.
The agent loop belongs to the backend CLI
(claude -p, codex exec). Cadora drives it, gates it, prices it —
which is exactly what makes its verdicts neutral and its cost ledger cross-vendor.
One class per backend. The same workflow runs on Claude or Codex and A/B-compares directly — design on one, code on another, measured to the node.
cadora doctorcheck your backend CLIs are on a tested version — offline, no model callscadora run <topology> --vision vision.md --cwd ./demodrive the gated workflow autonomously on your chosen backendcadora report <run-id>package the run as a portable, checksummed evidence packcadora eval <run-id>a deterministic pass/fail verdict with a CI-friendly exit codecadora compare <run-a> <run-b>the measured A/B — outcome, model, and cost, per nodeA self-contained
report.html + report.json + a SHA-256 checksums.txt
over every archived file. Mail it, print it, attach it to an invoice — verify it anywhere
with shasum -a 256 -c.
Gate verdicts (incl. a suite that ran zero tests), tamper findings, the human-review trail, and per-node cost by backend. It claims exactly what the archive recorded — and that it is checksummed, not signed.
The evidence isn't a summary Cadora writes about the run. The evidence is the run — captured, hashed, and handed over.
A real cadora compare on one topology: both backends
passed, and the cost came in at $1.98 vs $5.74 — measured,
reproducible, per node. The old "model X is always cheaper" heuristics don't survive contact
with a run archive.
Cadora re-runs the build/tests itself. A
runner that executes zero tests is vacuous and blocks
the run — it verified nothing. A missing toolchain is blocked_prerequisite,
not a fake failure.
cadora integrity catches
generated packages that impersonate real tools, unrecognized build substitutions, and tests
run against a swapped environment — the documented ways agents game their own checks.
The point is simple: Cadora verifies the boring facts itself. Modes:
audit (record), enforce (block), repair (one
constrained fix, then re-verify).
Reports dollars directly. Subscription-funded by default; metered API is an explicit opt-in.
Reports tokens; Cadora prices them from the public rate
table and flags the figure est.
Reports subscription credits — shown as credits, never pretended into dollars.
cadora usage and the dashboard's FinOps panel split cost by model,
backend, and funding — the same total everywhere, even when design ran on Claude and
code ran on Codex.
An autonomous run drives the agent with --dangerously-skip-permissions
inside the --cwd you give it — it can read, write, and run commands there with
your permissions.
A fresh directory, a git worktree, or a container — not your home dir, not a repo with secrets.
Every run prints a blast-radius banner and asks
once; CI bypasses with --yes.
Dashboard and MCP bind localhost only, no auth; Cadora refuses a public bind without an explicit flag.
pip install cadora · Python 3.10+, plus an authenticated claude or codex CLIcadora doctor · confirm at least one backend is readycadora run examples/aidlc.topology.yaml --vision examples/hackathon-hello.vision.md --cwd ./demo · a fun build in ~2 minutescadora report <run-id> · open the pack, then cadora compare a second backendFull walkthrough: docs/hackathon-quickstart.md · presenter's script: docs/demo-script-5min.md · docs/getting-started.md · github.com/yeychenne/cadora. Bring a throwaway repo and a backend login — you'll have a checked, priced build before the demo ends.