A second opinion for your
coding & finance agent.
The TruVerifAI plugin gives your AI agent a four-model blind-spot review — across coding (with proactive gates that prompt the right check before a risky commit ships) and finance (valuations, forecasts, capital allocation, credit, trade ideas). One model can be confidently wrong. Four, reasoning against each other, catch what one rubber-stamps.
# add the marketplace, install the plugin, reload /plugin marketplace add https://github.com/TruVerifAI/claude-plugins.git /plugin install panel-review@truverifai /reload-plugins # paste your tvai_… key into /plugin, then verify /panel-review:setup
The panel
Three ways to ask four models
Three primitives, one credit pool. Reach for synthesize when the question is small and reversible, deliberate when there's more than one defensible answer, and audit when you've already drafted something and want it stress-tested before it ships.
synthesize
quick checkA fast four-model second opinion on a bounded, reversible question — the idiomatic pattern, the conventional library, “is there a standard way to do this?” Returns an answer status, severity-tagged findings, and an advisory take.
deliberate
open decisionAn open-ended design decision with conflict-targeted revision. The four models reason independently; their disagreements are routed back as specific points each must defend or revise — then synthesized into a recommendation, not an averaged mush.
audit
stress-test a draftStress-test something you've already drafted — a diff before merge, a design doc, a migration plan, a security assessment. Four models hunt for blind spots, then revise their critiques against each other's. Returns a verdict + findings.
Each primitive ships in two profiles — coding and financial. The financial profile serves the full loop (generate / decide / verify) across any finance decision — valuation, forecasting, capital allocation, credit, and trading — weighing the affirmative case, not just downside. Information / critique, not advice.
The differentiator
Gates that prompt before you commit
Skills tell an agent what it could do; they're easy to skip. The gates make it deterministic — a PreToolUse hook intercepts the risky moment and asks for the right call, turning “you should review this” into “please audit this now.”
Audit before a risky commit
Before a git commit, the plugin classifies the staged diff locally — auth, secrets, billing, migrations, infra-as-code, a removed permission or validation check. If it's risky, the commit pauses and the agent is prompted to run audit. A PASS releases it.
Deliberate before a risky design write
Before a Write or Edit that sets a schema, migration, dependency, or trust boundary, the deliberate gate prompts a real deliberation. Tiered by default: it blocks only the high-confidence, hard-to-reverse forks; lower-confidence changes just get a nudge.
Privacy-clean & fail-open
The classifier runs on your machine; the backend only ever sees a repo fingerprint and hunk hashes — never your source, paths, or diffs. And every gate fails open: missing key, no network, our server down — it releases rather than trapping your agent. It never deadlocks.
Release it your way
A real audit/deliberate PASS clears the gate. Genuinely a false positive or already reviewed? Log a one-line skip (free) and move on — every skip is recorded. You stay in control; the gate is cooperative, not a wall.
Configure it in /plugin: turn gates on or off, set the deliberate mode to tiered, block, or advisory. The gate is a cooperative prompt layer, not a bypass-proof control. Gates are for the coding profile (they intercept a git commit or a Write); the financialprofile has no automatic gate — it's invoked through its skills, with a consumer-wired execution-boundary check on the roadmap.
Decision-grade output
A verdict, not a vibe
Every call returns a structured assessment your agent (or your dashboard) can branch on — not a wall of prose to re-read.
A headline verdict
audit returns approve / approve-with-caveats / request-changes / reject (financial: sound / sound-with-caveats / reconsider / reject). deliberate returns a clear / qualified / split recommendation. synthesize returns whether the answer is settled or contested.
Severity-tagged findings
A uniform findings list — each tagged critical, major, minor, or preference — so the things that matter sort to the top instead of hiding in paragraph four.
A recommended action
A derived action — coding: proceed / proceed-with-caveats / request-changes / escalate; financial: proceed / review-assumptions / gather-more-data / escalate — computed from the verdict and the worst finding, never from raw model agreement. A unanimous “reject” can't read as “proceed.”
Auto-activating skills
Eight skills, across coding and finance
In Claude Code, the skills load and activate at decision moments — no remembering to call anything. Three per profile (audit / deliberate / synthesize) plus two shared. Each is a plain Markdown skill you can read, and invoke by hand too.
audit-before-commit
Fires before you commit a hard-to-undo or security-boundary change — and on a soft “mind taking a look?” Routes the diff to audit.
deliberate-before-implementing
Fires when a decision has more than one defensible answer — schema, API shape, library, caching, concurrency. Routes it to deliberate.
synthesize-quick-check
Fires for a quick sanity check or “what's the idiomatic way?” — a fast four-model take in ~15-30s.
audit-financial
Fires before you act on a drafted finance decision, or to review a model/document — a backtest, valuation (DCF/LBO), forecast, risk report, capital-allocation memo, or credit approval. Routes to audit_financial. Information, not advice.
deliberate-financial
Fires when a finance decision has more than one defensible answer — which trade/investment, sizing, a valuation, capital allocation, a forecast assumption, a credit or accounting call. Routes to deliberate_financial.
synthesize-financial
Fires for a fast finance read or to generate candidate options — trade ideas, capital-allocation alternatives, valuation approaches, budget scenarios (each with its disconfirming case + a falsifiable test). Routes to synthesize_financial.
record-outcome-after-acting
Fires after the agent acts on a result, reporting back whether the second opinion actually changed the decision. Free.
skip-gate-when-not-needed
Fires when a gate is a genuine false positive or already-reviewed — logs a one-line skip so the change proceeds. Free, and every skip is recorded.
Outside Claude Code (Cursor, Codex CLI, Goose, any MCP client), the tools work the same — point the client at the MCP server and call them directly or from your rules file.
Pricing
Same credits as everything else
No separate MCP plan. Panel calls draw from your existing TruVerifAI credit pool — and the tools that just report back are free.
One credit pool
synthesize, deliberate, and audit deduct from the same monthly credits as the chat app and the REST API. No second invoice, no new pricing to learn.
Reporting back is free
record_outcome and record_gate_skip never cost credits. Logging a skip or telling us whether a review changed your mind is always free.
Burst limits built in
Per-key rate limits keep a runaway agent from blowing your credits: deliberate and audit at 15/min, synthesize at 40/min, up to 1,000 panel calls a day.
Quickstart
Installed in under a minute
Already have a TruVerifAI account? In Claude Code you're three commands from a four-model reviewer in your editor — and Cursor, Codex CLI, or any MCP client can point at the same server (cross-platform installer in the plugin repo).
Get an API key
In Settings → API, create a key and accept the AUP. The tvai_… secret is shown once — copy it.
Add the marketplace & install
Run /plugin marketplace add … then /plugin install panel-review@truverifai, then /reload-plugins.
Paste your key
Open /plugin → AI Panel Review → api_token, paste the key, and /reload-plugins to activate.
Verify & go
Run /panel-review:setup to confirm connectivity and loaded skills. Make a risky commit and watch the gate prompt an audit.
Give your agent a panel
Install the plugin, paste your key, and let four models catch the blind spots before they ship. Free to start — 50 credits, no card.