One repeated agent mistake. One configured gate.
ThumbGate checks proposed tool calls before execution in supported hook integrations. The $499 Setup Sprint turns one recurring failure in a supported local workflow into a configured gate, a regression test, and rollout and rollback evidence.
Illustrative example, not a recorded result. Matching destructive actions warn by default; advisory integrations do not guarantee execution blocking.
Free (MIT): npx thumbgate init
Claude Code, Cursor, Codex, or similar agents.
Dashboard: npx thumbgate dashboard --open · /thumbgate-dashboard · thumbgate-dashboard · demo · strict-mode deny example ↓
Thumbs teach. The gate enforces.
Pre-action checks—and the systems that refine them.
Corrections become local lessons; repeated negatives become gates; the next action is checked before execution. Click a step for the under-the-hood demo.
Capture feedback
Record explicit 👍 or 👎 with the action and outcome context.
Under the hood ↓Remember locally
Store a reviewable lesson that survives sessions without writing into model weights.
Under the hood ↓Rank and refine
Re-rank lessons for the action. Repeated negatives can promote to blocking gates; stale gates expire.
Under the hood ↓Gate the next action
The action is allowed, warned, or denied before the tool proceeds.
Capture feedback
Each reviewed outcome closes the loop—under your control. Lessons are re-ranked per action, repeated failures can promote into gates, and stale auto-promoted gates expire. The firewall improves from explicit feedback without retraining the model or silently rewriting policy.
A managed gate for one painful workflow.
Hard deny where configured. Honest warning everywhere else.
proposed git push --force origin main mode strict enforcement rule protect-main source core protection · strict mode decision DENY reason force-push to protected branch next decision recorded before execution
Why not a static hook? Prompts compact away; static bash hooks don't learn from 👍/👎. Detected secret exfiltration and gate-process bypass attempts are denied by default. Matching destructive actions warn by default and deny in strict mode. This is a managed strict-mode example, not a claim that every free install blocks every risky command automatically.
Read the test-backed verification evidence →White paper · Scorecard · Architecture · Dogfood, not customer stories
Enforcement by agent and mode
| Agent | Enforcement | Default (secrets) | Default (destructive) | Strict mode |
|---|---|---|---|---|
| Claude Code, Codex, Gemini CLI | Blocks before tool runs | DENY | WARN | DENY |
| Cursor, Cline, OpenCode | Advisory | DENY | WARN | DENY |
| Amp | Feedback only | Logged | Logged | Logged |
Five questions, answered plainly.
npm pack thumbgate and read it yourself. Your lessons live in a local SQLite file you own, not a server you can't query. The $499 Setup Sprint adds a done-with-you failure map, one configured gate, and its regression proof, not access to intelligence withheld from the free install.What this does not do.
- Does not retrain model weights or silently rewrite production policy.
- Does not contain the blast radius — that is a sandbox's job.
- Default install warns on destructive actions; hard deny needs strict mode (
THUMBGATE_STRICT_ENFORCEMENT=1). - The $499 Setup Sprint covers one supported local workflow. Multi-system implementation, SSO, a hosted audit log, and compliance packaging require separate scope and are not generally available.
Block the mistake your agent made last week.
Run npx thumbgate init (MIT), or book a $499 setup.