Rextor, in plain words
Step 01
Install the app. It joins your repo like a reviewer.
GitHub App · two-click install
Step 02
Every PR that touches money-code gets audited — machines find, citations prove, forks test.
diff-scoped · Slither · Foundry fork-sim
Step 03
The verdict lands on-chain — permanent, checkable by anyone, impossible to quietly edit.
attestation · Tempo testnet + HyperEVM mainnet
How a review runs
1.0
Scope the diff
The PR diff becomes a change-set of functions and callsites — not a vibe read of the whole repo. New external calls, permission changes, and trust‑boundary crossings get flagged for the passes below.
2.0
Detect deterministically
Slither runs in a network‑isolated, read‑only container. Tool errors produce an explicit INCOMPLETE report — never a fabricated clean pass.
- analyzer pass · sandboxed
- runtime
- Slither 0.11.6, Docker --network none --cap-drop ALL
- scope
- diff-scoped: changed functions, new callsites, trust-boundary deltas
3.0
Triage with citations
An LLM constrained to three cited operations: reclassify a severity, dedup into an equal‑or‑higher one, or add a finding — only with a cited line range. It assigns no score, and any uncited operation is discarded and counted.
- triage policy · temperature 0
- allowed
- reclassify (original preserved) · dedup (never escalates) · add (cited lines only)
- forbidden
- scoring · inventing findings · operating without a citation
4.0
Prove by execution
Criticals are reproduced on a fork of the target chain — Foundry against live state. A finding that survives the fork is confirmed; one that doesn't is reported as unproven. Reading code is opinion. Running it is evidence.
5.0
Anchor the verdict
The verdict lands on‑chain: commit hash, findings hash, risk score, agent identity. The score comes from a published rubric — anyone can recompute it from the findings alone, then check it against the chain.
- attestation · real record
- reviewId
- keccak256("rextor/review/v1|" + owner/repo + "|" + pr + "|" + headSha)
- findingsHash
- sha256 of the canonical findings JSON — printed in every review comment
- score
- rubric: critical 60 · high 25 · medium 10 · low 3, cap 100
- agent
- rextor-audit[bot] · 0x5f2b…7f37
- chains
- Tempo 42431 ↗ · HyperEVM 999 ↗
What actually ships
Live rows carry their receipt — every green check links its artifact. Competitor marks are datestamped, not guessed.
Live now · 15 receipts
| Capability | Generic AI review | One-shot audit bots | Rextor Audit |
|---|---|---|---|
| Built for money‑code — analyzer grounding | stage 2.0 | ||
| Every PR an audit event (diff‑scoped, continuous) | stage 1.0 | ||
| Deterministic, recomputable risk score | rubric | ||
| Execution‑proof findings (fork‑sim PoC) | stage 4.0 | ||
| Verdict on‑chain, native to the target chain | Tempo · live · HyperEVM mainnet · live · Solana devnet · live | ||
| Integrity rails — INCOMPLETE never silent | stage 3.0 | ||
| Suggested fix per finding (reviewable diff) | smoke PR · suggested diff | ||
| Repo config + severity merge gate | smoke PR · severity gate failure on high finding (App-authored check-run) | ||
| Durable full report (IPFS) + anyone‑can‑verify | smoke PR · CID + on-chain hash | ||
| Verifiable agent identity (ERC‑8004) | agentId 50891 · mainnet · live | ||
| Public agent reputation ledger | automated ingestion pipeline · evidence-backed (findingsURI+hash) · broadcast config-gated | ||
| Cross‑PR memory (dismissals) | smoke PR · learnings ledger | ||
| Chat in PR | smoke PR · bot reply | ||
| Chain cost‑model reasoning | real receipts: Tempo 0.0125–0.154 (6 txs) · HyperEVM mainnet 0.0000468 HYPE (verify=true) | ||
| Living dashboard — ledger, history, identity | live dashboard |
shipped · partial · not evidenced — competitor marks relative to public product pages as of 2026‑09‑25.
Rider tracks — Ethereum (1), Base (8453), Arbitrum One (42161), Robinhood Chain (params at integration) — anchor via the home-chain attestation; targetChainId recorded per review (foundry hint).
Where it runs
Tempo flagship · chainId 42431
Full loop, live. Slither, fork‑sim, attestation v2 deploy #3 — 27 reviews anchored — contract ↗
HyperEVM primary · chainId 999
Mainnet, live. Attestation v2 deployed 2026‑09‑21, agent registered active — first mainnet attestations pending — contract ↗
Riders config‑level
Ethereum 1 · Base 8453 · Arbitrum 42161 · Robinhood Chain — deterministic pass; verdicts anchor home‑chain with a target‑chain field. No per‑rider code paths.
What Rextor can't do
The integrity rails, in plain words — the refusals are the product.
It can't invent findings.
Every operation carries a cited line range or it is discarded. The agent never scores anything — the published rubric does, and anyone can recompute it.
It can't fake a clean pass.
A broken analyzer produces status = 1, attested on-chain — a riskScore: 0 can never masquerade as "all clear".
It can't be hijacked by PR content.
Titles, comments, and code are adversarial by assumption — rendered inert, prompt-injection hardened, simulation sandboxed.
It can't lose money.
The attestation key has no fund path. Worst case from key theft: visible, attributable bogus attestations. There is nothing to steal.
Questions, answered straight
What does it cost?
Free. If that ever changes, pricing ships on this page first — no surprise invoices.
Which chains does it run on?
Tempo testnet runs the full loop today, and verdicts anchor on HyperEVM mainnet as of 2026‑09‑21. Ethereum, Base, Arbitrum, and Robinhood Chain ride the same deterministic pass with home‑chain anchoring. Solana devnet is live ↗.
What is a "verdict on-chain"?
Every review ends with a small record written to the blockchain — commit hash, findings hash, risk score, agent identity. After it lands, nobody can edit it, including us. See one ↗
Does this replace my auditor?
No. Rextor makes every pull request an audit event between the human audits. Your annual engagement still matters — this is the continuous layer underneath it.
What happens when the tooling breaks?
The run marks itself INCOMPLETE and attests that on-chain. A broken analyzer can never look like a clean pass — failure is a visible, checkable state.
Can I see it working before installing?
Yes — live verdicts on Tempo ↗, and the whole system is built in the open: public repo ↗.
Install once. Every PR after that is an audit event.
Two‑click install on GitHub. The agent reviews the PRs that touch money‑code and anchors each verdict on‑chain.
Install the GitHub App