๐Ÿ‘๐Ÿ‘Ž
โ— Human-in-the-Loop Enforcement for AI Agents

AI agents repeat mistakes.
Yours won't.

๐Ÿ‘Ž blocks dangerous and dumb mistakes before they run
๐Ÿ‘ reinforces what worked across sessions

For power users of Claude Code, Cursor, Codex, Gemini, Amp, and any MCP agent who are tired of fixing the same mistakes every session.

Every mistake makes your agent stronger. ThumbGate turns thumbs-up/down feedback into hard gates that block bad actions before they execute. Think of it as an immune system for your AI agent โ€” it learns from your corrections, distills vague signals from conversation context, and enforces lessons automatically. More errors = stronger gates. The safety net for vibe coding that gets more valuable the more chaos it encounters.

$ npx thumbgate init click to copy

One gateway across the agent surfaces you already use

Try the dashboard before you buy anything

๐Ÿ” Live Dashboard Demo

Search lessons, inspect gates, see blocked actions โ€” all with real sample data. No signup, no install required.

Open Dashboard Demo โ†’

โ›” Gate Reasoning Chains

Every block explains WHY: which pattern matched, what evidence triggered it, whether it was auto-promoted from your feedback or manually set.

๐Ÿ“Š Org Dashboard (Pro)

See what ALL your agents are doing. Adherence rates, risk alerts, top blocked gates โ€” across every session in your org.

๐Ÿงช Thompson Sampling

Confidence tiers (none/low/medium/high) tell you when to trust the model vs fall back to rules. No guessing.

๐Ÿชž History-Aware Lessons

When a thumbs signal is vague, ThumbGate can reuse the last ~10 messages plus the failed tool call, propose what went wrong, and turn that into one linked lesson instead of a dead-end note.

How ThumbGate compares and when to use it

Three steps to mistake-proof vibe coding

1

Feedback

Give ๐Ÿ‘ or ๐Ÿ‘Ž on your AI agent's actions. Feedback is stored in a SQLite+FTS5 lesson DB. If the signal is vague, ThumbGate can distill it from the recent conversation window and failed tool call before promotion. Example: you ๐Ÿ‘Ž a risky migration โ†’ it auto-promotes to a "never run DROP on prod" gate.

2

Distill + Rules

Repeated failures auto-promote into prevention rules. Thompson Sampling adapts which rules fire, and the reflector lane can propose a reusable rule from the same transcript so high-risk patterns get stricter enforcement while low-risk ones stay relaxed.

3

Gates

Rules become Pre-Action Gates that block your agent before it repeats the same mistake. Your agent can't force-push, skip tests, or repeat a refactor you already rejected. No more fix-loops.

Built to get stronger from chaos, not break from it

Errors are fuel

Every agent mistake becomes a prevention rule. Every thumbs-down becomes a gate. The more your agents fail, the more ThumbGate learns. Fragile systems break under chaos โ€” ThumbGate feeds on it.

Organizational learning

One developer's thumbs-down becomes the whole team's protection. Lessons compound across agents, sessions, and projects. Your 100th agent inherits everything the first 99 learned.

Enterprise-grade governance

CISOs fear agent chaos. ThumbGate turns that chaos into structured audit trails, per-agent governance scores, PII scanning, and credential attestation. Compliance officers get a single dashboard, not scattered logs.

Every guardrail pattern experts recommend, built in

Don't trust โ€” verify

Gate reasoning chains show exactly WHY an action was blocked: pattern match, evidence, fire count. No hidden decisions.

Real tools, not wishes

MCP tools call real enforcement โ€” gates physically block tool calls, not ask the agent nicely. External system, not prompt tricks.

Force models to show work

Every gate decision includes step-by-step reasoning. Thompson Sampling confidence tiers tell you when to trust the model vs fall back to rules.

Log everything, learn automatically

Audit trail records every allow/deny/warn. Repeated failures auto-promote to gates. Org dashboard shows all agents across your team.

Keep one correction thread

Feedback sessions let you open, append, and finalize one linked correction instead of spraying follow-up notes across separate records. Better lessons, less operator friction.

AI Control Tower: monitor, alert, and act

Domain skill packs

Installable best-practice rules for Stripe, Railway, database migrations. Auto-match by task context. 100% eval pass rate with packs vs 0% without.

Tool-level KPIs

P50/P90/P95 latency and success rate per tool. Drill from gateway health down to the exact tool causing issues.

SLO alerts

Define thresholds for success rate and latency. Breach triggers webhook alerts to Slack, Teams, or Discord automatically.

Access anomaly detection

Tracks authorized vs failed access. Moving-average spike detection flags per-agent abuse. Security governance on autopilot.

Metered outcome pricing

$0.10 per blocked mistake (Pro). Minimum $19/mo floor. You pay for value delivered, not seats warmed.

Ambient daily digest

Push Slack/Teams/Discord summary at 9am. Weekly stats export for build-in-public content. Check in once a day, not babysit.

82% fewer tokens with progressive skill loading

3-tier disclosure

L1 metadata at startup (27 tokens). L2 rules on activation. L3 reference files on demand. Agents load only what they need.

Skill factory

Agent auto-generates new skill packs from recurring failures. Lessons become NEVER/ALWAYS rules. No manual curation needed.

Security hardening

Credential attestation per tool call. Privilege escalation detection against MCP profiles. Dependency gate blocks supply chain attacks.

Your agents write 57% of PRs. ThumbGate makes sure they don't repeat mistakes.

Run governance gate

Before a background agent opens a PR, check its failure history. Agents with >50% fail rate get blocked. Protected branches get extra scrutiny.

Auto-capture from CI

CI pass/fail auto-generates structured feedback. No human in the loop. The agent learns from every merge and every rejection.

Governance report

Per-agent pass rates, gate blocks, run types, blast radius warnings. Know which agents are reliable and which need guardrails.

terminal
# Install in 30 seconds
npx thumbgate init
 
# Your agent tries to delete production db...
โ›” Gate blocked: "Never run DROP on production tables"
# Rule auto-generated from your previous ๐Ÿ‘Ž
 
# Works with your existing setup
npx thumbgate init # bootstraps .thumbgate and .mcp.json
# Use repo guides for Claude plugin, Codex app plugin, Cursor plugin, and OpenCode profiles

Start free. Try Pro for 7 days. Cancel anytime.

Free
$0
Forever free ยท Local-only, single dev

For solo devs experimenting with gates on one machine.

  • Unlimited feedback captures ยท 5 lesson searches/day ยท unlimited recall
  • 5 auto-promoted gates plus the core safety policy
  • All MCP integrations (Claude Code, Cursor, Codex, etc.)
  • PreToolUse hook blocking
  • Local SQLite lesson DB
  • Setup guide for all agents โ†’
Install Free
Pro
$19/mo
or $149/yr (save 35%) ยท Personal dashboard + exports

For individual operators who want a personal local dashboard with proof-ready exports.

Go Pro when: you've blocked 20+ actions, want your own dashboard, or need DPO exports without extra setup.
What your Pro dashboard looks like
โœ… gate:no-force-push โ€” blocked 12 times
โœ… gate:require-tests โ€” blocked 8 times
โŒ gate:no-drop-prod โ€” FIRED (blocked DROP TABLE)
DPO pairs exported: 47 | Lessons: 23 active
  • Everything in Free, plus:
  • Visual gate debugger โ†’ see every blocked action and the gate that fired so you can trust the system in minutes
  • Auto-connect โ€” activate once with your license key, then your running agents appear automatically on your local dashboard
  • DPO training data export โ†’ turn real thumbs-downs into ready-to-use preference pairs for fine-tuning (LoRA / JSONL)
  • Model Hardening Advisor โ€” get recommendations on when and how to fine-tune your model to natively avoid recurring failures
  • Personal local dashboard โ€” every Pro user gets their own localhost dashboard without extra cloud setup
  • Founder-license support โ€” we help you wire your riskiest flows (migrations, force-pushes, CI) in the first week
7-DAY FREE TRIAL

Founder pricing โ€” preserved $49 one-time founder link for early buyers who still want that path.

Mem0 Pro: $249/mo. ThumbGate Pro: $19/mo.

Team
$12/seat/mo
Starts at $36/mo for 3 seats ยท or $297/yr ยท Shared lessons + org visibility

For teams running multiple agents across shared repos who need one lesson to protect everyone.

Start Team when: you need a shared lesson DB, org-wide metrics, or workflow hardening that survives handoffs.
  • Hosted shared lesson DB โ€” one dev's ๐Ÿ‘Ž on a bad migration protects every agent on the team
  • Org dashboard โ€” active agents, gate hit rates, risk agents, and proof-backed team metrics in one place
  • Generated hosted review views โ€” constrained cards, lists, and callouts assemble team, incident, and rollout dashboards without custom page work
  • Gate template library โ€” pre-built guardrails for force-pushes, skipped tests, destructive SQL, and evidence-before-done
  • Workflow hardening sprint โ€” intake, rollout plan, and evidence pack for the first shared workflow
  • Product Hunt and campaign attribution stay tied to the team funnel instead of generic referral traffic
Start Team Rollout

Hosted team rollout starts with the workflow sprint intake so we scope the shared workflow honestly.

Start the Team lane with one workflow that matters

The recurring value is shared enforcement. Start with one workflow, one owner, and one blocker. The hosted intake gives us a proof-backed path into the shared lesson DB, org dashboard, and team-wide gate rollout.

Common questions

Does ThumbGate support model fine-tuning?
Yes. ThumbGate Pro includes a Model Hardening Advisor and LoRA JSONL export. Pro users can export their episodic memory as DPO (Direct Preference Optimization) pairs to fine-tune local models (like Llama 3 or Mistral) so they natively avoid repeating known mistakes.
ThumbGate's intelligence is context, not weights. It doesn't touch the model โ€” it injects past feedback into context so your agent is conditioned by your corrections. Think of it as a behavioral immune system, not a training pipeline. The gate blocks are hard enforcement, not soft suggestions.
SQLite+FTS5 lesson DB for fast full-text search. MemAlign-inspired dual recall (principle-based rules + episodic context). Thompson Sampling for adaptive gate sensitivity per failure domain. LanceDB + Apache Arrow for local vector search with Hugging Face embeddings. ContextFS for context assembly. Bayesian belief updates on each memory. PreToolUse hook enforcement blocks known-bad actions before execution. All local-first โ€” no cloud required.
ThumbGate works with Claude Code, Cursor, Codex, Gemini CLI, Amp, OpenCode, and any other MCP-compatible agent. Cursor ships with a plugin bundle in this repo. Codex ships with a repo-local app plugin profile. VS Code works when you run an MCP-compatible agent inside it, but this repo does not ship a standalone VS Code extension today.
No. Free keeps local enforcement on your machine with 5 daily feedback captures, 10 lesson searches, unlimited recall, gates, and hook blocking. No cloud account is required. Pro adds a personal local dashboard on your machine plus DPO export. Team is the hosted rollout lane when your group wants a shared lesson database, generated hosted review views, org dashboard, and team-wide gate visibility.
ThumbGate can reuse the last ~10 messages, the failed tool call, and later follow-up notes to propose whatWentWrong, whatToChange, and a reusable lesson. The feedback session stays linked instead of creating duplicate records.
Prompt rules are suggestions that agents can ignore. Pre-Action Gates are enforcement โ€” they block the action before execution. Gates are auto-generated from your feedback and use Thompson Sampling to adapt over time.
Pro is $19/mo or $149/yr (save 35%). Team starts at $36/mo for 3 seats, then $12/seat/mo, or $297/yr at the 3-seat floor. Cancel anytime. No contracts.

Get notified when we ship

Release notes, new gate patterns, and agent governance insights. No spam. Unsubscribe anytime.

Ready to stop fixing the same AI mistakes?

Try the dashboard demo first. Install in 30 seconds. No account required.

Try Dashboard Demo Try Pro free for 7 days โ†’ $19/mo