High-stakes AI decisions,
vetted before they ship
Panel Review runs your agents' highest-stakes work, the risky diffs and the design, architecture, and security calls, through four frontier models that argue it out, and blocks what does not hold up before it ships. Run it on your own keys or your own models, keep every line of code inside your environment, and store nothing but usage metadata if you choose. Evaluate the whole thing in a disposable sandbox that never touches your real systems.
Your code stays in your tenancy. Run on your own keys or models. Nothing stored unless you choose.
Intro
Independent review that blocks risky changes before they ship
No single reviewer catches everything: people cannot keep pace with AI-generated code, and any one model has blind spots. Panel Review puts four frontier models from four vendors on the riskiest changes and design decisions, makes them argue it out, and holds them at the write or commit until they pass.
Cross-vendor blind-spot review
Four frontier models from four different providers review a change independently, then conflict-target each other's positions, and return one decision-grade verdict with severity-tagged findings. Not four opinions to reconcile yourself.
It blocks before the merge, not after
Hook code next to your agent intercepts risky writes and commits and routes them to the panel. Auth, secrets, money, migrations, a removed guard, or your own critical code (custom floors) do not ship unreviewed. The gate blocks; the panel decides; you proceed with evidence.
Proven on our own production code
Ten weeks, 479 review calls building TruVerifAI with the gates armed: 66% of reviews changed the agent's decision, 76% on security, under 1% of audits passed clean, and the panel escalated to a human 151 times.
Works on every surface your team uses
Claude Code, Codex, Cursor, Copilot and VS Code, Gemini, Antigravity, plus a git pre-commit hook. The same review everywhere, armed by one command.
Security
Your code stays in your tenancy
Send your security team straight here: truverif.ai/enterprise#security. Every claim is verifiable against our public client source.
Only the diff you submit ever leaves
A gate check sends a hashed repo fingerprint, content hashes of the changed hunks, and the local classifier's category labels. Not your source, not your file paths. A panel review sends only the diff or context your agent explicitly submits.
You choose what is stored
Persistence is on by default (a dashboard audit trail of your eval); turn it off and review content is not stored our side, only usage metadata. You can delete stored content and revoke access at any time.
Run it on your own models
Bring your own model keys (BYOK), or run the whole panel on your own models in your own AWS Bedrock or GCP Vertex tenancy (BYOM). Review content is then processed by the models you already operate and never reaches our accounts.
Self-minted, revocable keys; fail open
Sign-in is a browser device flow (the CLI never sees a password). The key is revocable server-side and lives only on the disposable machine. If our server is unreachable, your work proceeds with a visible notice: the gates never block you on our outage.
Delete the machine, delete everything
The sandbox runs on a GitHub Codespace or your own VM, in your tenancy. Deleting it removes the machine and everything on it. Uninstall is the machine ceasing to exist.
Open source, so you can verify it
Everything this command installs is open source: MIT-licensed, zero runtime dependencies. Read every line at github.com/TruVerifAI/init before you trust it.
Set up
Stand it up, then watch a gate fire
Everything runs on a machine in your own tenancy. Nothing installs on a real laptop, and deleting the machine leaves nothing behind. Budget about fifteen minutes end to end.
- 1
Open a disposable machine in your tenancy
On the template repo, click Use this template to make your own copy, then open it as a GitHub Codespace. Prefer AWS? Any machine with Node 18+, Python 3, and git works (an EC2 box or your own VM): clone your copy and use its terminal.
- 2
Install your coding agent and sign in
The gates arm on whatever agents are already on the machine, so install the one your team uses and sign in with your own account, for example
npm install -g @anthropic-ai/claude-codethenclaude. Claude Code, Codex, Cursor, Copilot, Gemini, and Antigravity all work the same way. - 3
Arm the review gates
Run
npx @truverifai/init. It signs you in to TruVerifAI by device flow (approve in your own browser, no shared credential) and arms the write and commit gates plus a git pre-commit gate. Your first login gets free evaluation credits. Verify withnpx @truverifai/init doctor; the full per-agent setup lives at truverif.ai/settings/mcp. - 4
Watch a gate block a risky change
The repo ships throwaway demo code fenced as a custom floor. Run
npm run demo:risky-change, then commit it: the commit gate stops it, cites the trade-core floor, and routes it to a four-model review. That is the product in one motion. - 5
Bring your own code and floors
Drop any non-critical repo of yours onto the machine and build on it for days. Mark the files that matter with
npx @truverifai/init floors init(or just ask your agent), and confirm coverage withfloors check --preview. - 6
Keep your data where you want it
Turn content persistence off at /settings/mcp, run the review on your own keys (BYOK), or run the whole panel on your own models in your own cloud (BYOM) so review content never reaches our accounts.
- 7
Tear it down
npx @truverifai/init uninstallremoves the gates and revokes your key server-side, then delete the Codespace or VM. Uninstall is the machine ceasing to exist.
Cost is compute only, often $0 on the Codespaces free tier, single digit to low tens of dollars for a multi-day eval. Review credits are on us.
Try it on code you understand
Spin up the sandbox, point it at a non-critical repo of your own, and see what the panel catches. It costs you a browser tab and a few minutes.
Launch the sandbox