What is the ThumbGate tech stack?

SQLite+FTS5 lesson DB, MemAlign-inspired dual recall, Thompson Sampling for adaptive checks, LanceDB vector search with Hugging Face embeddings, ContextFS context assembly, Bayesian belief updates, and PreToolUse hook enforcement.

What AI agents does ThumbGate work with?

Claude Code, Cursor, Codex, Gemini CLI, Amp, Cline, OpenCode, and any MCP-compatible agent.

How does ThumbGate reduce host blast radius for high-risk local runs?

ThumbGate combines pre-action checks with execution guidance. Workflow Sentinel predicts risky local actions before they execute, and high-risk runs can be routed into Docker Sandboxes instead of running directly on the host. Team workflows also have a signed hosted sandbox lane for isolated automation dispatch.

How does ThumbGate prevent AI slop and protect brand authenticity?

AI slop happens when agents act without human judgment as a hard check — generating repetitive, generic outputs that erode trust and dilute your brand. ThumbGate inserts human thumbs-up/down between AI intent and execution. Every thumbs-down becomes a prevention rule that blocks the bad pattern permanently. Every thumbs-up reinforces what 'good' looks like for your specific context. Your agent's outputs reflect your actual standards, not generic AI patterns. This is authenticity enforcement at the tool-call level.

How is ThumbGate different from SpecLock?

SpecLock requires manually writing constraints or compiling them from a PRD. ThumbGate learns automatically from thumbs-up/down feedback and auto-generates prevention rules from repeated failures. SpecLock locks files from modification; ThumbGate blocks specific actions (like force-push) before they execute.

How is ThumbGate different from Mem0?

Mem0 is cloud-hosted memory for AI apps. ThumbGate is local-first enforcement. Mem0 remembers context but cannot block actions. ThumbGate captures feedback, promotes it to prevention rules, and physically blocks tool calls that match known failure patterns via PreToolUse hooks.

👍 👎 Pre-action gates for AI coding agents

Stop paying for the same AI mistake twice.

ThumbGate turns a single thumbs-down into a rule that blocks the next repeat before Claude Code, Cursor, Codex, Gemini CLI, Amp, Cline, or OpenCode runs the tool call.

$ agent attempts risky action

BLOCKED: never run DROP on production tables

Agent must choose a safer plan with evidence.

$ npx thumbgate init click to copy

Install Free CLI Get Pro — $19/mo

MIT open source Local-first Works with MCP-compatible agents Verification evidence in GitHub

First-Dollar Activation Path

Block your first repeated AI mistake in 5 minutes.

Install, give one thumbs-down, and let the PreToolUse hook stop that failure from rerunning. Upgrade only when you need dashboards, exports, or shared team enforcement.

1. Install ThumbGate

Run npx thumbgate init in your repo. Claude, Cursor, Codex, Gemini, Amp, Cline, and OpenCode are supported through local hook or MCP-compatible setup paths.

2. Give feedback

Use thumbs up for a good pattern and thumbs down for a failure. ThumbGate distills the correction into a concrete prevention rule.

3. The repeat gets blocked

The next matching tool call is stopped before execution, with a reason the agent can use to choose a safer plan.

👍👎

Thumbs-down once, blocked forever Block repeat hallucinations before the model sees them

For consultancies, platform teams, and AI product teams that want workflow governance, CLI-first rollout, and a reliable operator.

Have one AI-agent failure that keeps repeating? Start with one real workflow, one repeated failure pattern, enforceable pre-action gates, and a short audit note your team can keep.

Prove one blocked repeat before asking anyone to buy. Give thumbs up when the agent follows your standards, thumbs down when it misses. Upgrade after one real blocked repeat.

Workflow governance for isolated execution: ThumbGate pairs policy checks with Docker Sandboxes, signed hosted sandbox dispatch, Changeset evidence, and exact main-branch merge commit verification before release claims ship.

Claude Code · Claude Desktop · Claude Extension

The fastest path for Claude users: install the extension and start blocking mistakes.

ThumbGate ships a published Claude Desktop extension bundle (.mcpb) you can install today. Claude Code users can also add the repo marketplace plugin immediately. No waiting for directory approval.

1. Install for Claude Code

Run npx thumbgate init --agent claude-code or add via claude mcp add thumbgate -- npx --yes --package thumbgate thumbgate serve

2. Or install the Claude Extension

Download the .mcpb bundle for Claude Desktop, or use the repo marketplace: /plugin marketplace add IgorGanapolsky/ThumbGate

3. Give feedback, checks auto-generate

Type thumbs down when Claude makes a mistake. ThumbGate distills a lesson from up to 8 prior entries and blocks the pattern permanently via PreToolUse hooks.

Download Claude Extension (.mcpb) Claude Desktop setup guide Claude plugin docs

Claude Code Skill: Type /thumbgate in any Claude Code session. Auto-triggers on “check”, “feedback”, “block mistake”. Free skill on top of the same local gateway.

ChatGPT Entry Point · Live ThumbGate GPT for ChatGPT

Use the GPT as a preflight desk for risky commands, refunds, deploys, and PR actions.

ThumbGate should meet users where they already ask AI for help. The live GPT is the fastest way to preflight a risky action, capture a typed thumbs-up/down lesson, and prove the enforcement loop before installing anything. As ChatGPT ads roll out, this matters more: ChatGPT can stay the discovery and checkpointing layer, while ThumbGate remains the hard execution boundary after npx thumbgate init.

1. Open the live GPT

Paste a proposed command, file edit, merge, deploy, refund, invoice, or API call and ask whether to allow, block, or checkpoint it.

2. Save the typed signal

Reply in chat with thumbs up: or thumbs down: plus one concrete sentence. Do not rely on ChatGPT's native rating buttons for ThumbGate memory.

3. Enforce locally

Run npx thumbgate init in the repo so Pre-Action Checks block repeated mistakes before the coding agent executes them.

Open ThumbGate GPT ChatGPT Actions setup Why ChatGPT ads need checks

Find it fast: if the direct link does not open, go to Explore GPTs, search ThumbGate, and choose the GPT by Igor Ganapolsky in Programming. Plain English rule: ChatGPT is the discovery and memory surface for advice, checkpointing, and typed feedback capture. One typed signal becomes one remembered rule. The hard Reliability Gateway still runs in the local agent or CI lane.

Positioning

Enforcement is the missing layer in AI orchestration.

Big orchestration suites unify data, routes, and decisions. ThumbGate sits closer to the moment of execution: the point where an agent is about to run a shell command, ship a PR, approve a release, or repeat a mistake you already corrected. That is where workflow trust is won or lost.

Broad orchestration platforms

Good at customer journeys, routing, and cross-system context. Weak when you need a coding agent or automation to stop before a destructive or low-trust action actually runs.

ThumbGate

Turns operator feedback into Pre-Action Checks. It does not just remember the mistake. It blocks the repeat at the tool-call boundary across Claude Code, Cursor, Codex, Gemini, Amp, Cline, and OpenCode.

The stack that makes sense

Use orchestration to decide what should happen next. Use ThumbGate to decide what is allowed to execute. That is the control layer enterprises actually need once agents touch repos, terminals, CI, or production workflows.

Compare orchestration vs enforcement → Platform team rollout → Regulated workflow pattern →

Autoresearch Safety Pack

Stop self-improving coding loops from hacking the benchmark.

Autoresearch loops run experiments, inspect metrics, and accept better variants. ThumbGate gives those loops a Reliability Gateway: Pre-Action Checks for skipped holdout tests, fake proof, reward hacking, unsafe edits, and promotion without verification evidence.

Guard the metric

Require primary and holdout checks before an agent can call a variant better. Block cherry-picked runs and missing baselines.

Preserve proof trails

Promotion needs commands, logs, changed files, and verification evidence so the win survives review instead of becoming a vague claim.

Ship into CI

Start with templates for npm test, Playwright duration, bundle size, lint, and CI failures, then add Team checks for shared workflows.

Read the Autoresearch guide Start Pro

Current release

Three steps to stop repeated AI failures

Feedback

Give 👍 or 👎 on your AI agent's actions. Feedback is stored in a SQLite+FTS5 lesson DB. In the current Claude auto-capture hook, a vague thumbs-down can distill from up to 8 prior recorded entries and the failed tool call before promotion, then stay linked to a 60-second feedback session. Example: you 👎 a risky migration → it auto-promotes to a "never run DROP on prod" check.

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.

Checks

Rules become Pre-Action Checks 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.

Enforcement

Checks block. They don't ask nicely.

Don't trust — verify

Every block shows why: pattern match, evidence, confidence score.

Real tools, not wishes

Checks physically block tool calls. Not prompt tricks. Not "please don't."

Force models to show work

Reasoning chains on every check decision. Thompson Sampling confidence tiers.

Log everything, learn automatically

Repeated failures auto-promote to checks. Org dashboard shows all agents.

Keep risky runs off the host

When Workflow Sentinel predicts a risky local action, ThumbGate can recommend Docker Sandboxes before the agent touches the host filesystem or broader credentials.

Ship with versioned proof

Changesets, SemVer, sync checks, and verification evidence make new package releases inspectable before a buyer trusts the next rollout.

Install for Your Agent

Claude Code

npx thumbgate init --agent claude-code

Wires PreToolUse hooks automatically

Cursor

npx thumbgate init --agent cursor

4 skills: feedback, rules, search, recall

Codex

npx thumbgate init --agent codex

6 skills including adversarial review

Gemini CLI

npx thumbgate init --agent gemini

Gemini CLI integration

Amp

npx thumbgate init --agent amp

Amp agent integration

Any MCP Client

npx thumbgate serve

MCP stdio server for any compatible client

Claude Desktop

Add to your claude_desktop_config.json:

{
  "mcpServers": {
    "thumbgate": {
      "command": "npx",
      "args": ["--yes", "--package", "thumbgate", "thumbgate", "serve"]
    }
  }
}

Official directory review is separate. Claude Code users can install immediately with /plugin marketplace add IgorGanapolsky/ThumbGate and /plugin install thumbgate@thumbgate-marketplace.

Pricing

Stop paying for agent mistakes you already fixed.

Free

Block repeated mistakes daily. Forever free for solo devs.

Unlimited captures, 5 active rules. Enough to make ThumbGate part of your daily flow. Upgrade when you want the dashboard, exports, or unlimited rules.

Unlimited feedback captures — every thumbs-down, every session
Up to 5 active auto-promoted prevention rules
No recall or lesson search
No exports (DPO, Databricks, HuggingFace)
All MCP integrations (Claude Code, Cursor, Codex, Gemini, Amp, any MCP agent)
PreToolUse hook blocking with built-in safety checks (force-push, destructive SQL, secrets)
Setup guide for all agents →

$ npx thumbgate init click to copy

Install Free

Pro

$19/mo

Stop paying tokens to re-correct the same agent mistake across sessions.

Block every repeat mistake — unlimited feedback captures and prevention rules (Free caps at 5 active rules)
Never re-explain a correction — lesson recall and search across sessions on every agent surface
See exactly which rule fired — Visual check debugger for every blocked action and the check that fired
One install, every agent — Auto-connect so supported agents appear automatically after setup (Claude Code, Cursor, Codex, Gemini, Amp, Cline, OpenCode)
Fine-tune your local model on what your team actually wants — DPO training data export with ready-to-use preference pairs for fine-tuning
Audit-ready enforcement proof — Personal local dashboard for the individual operator with auditable block history
Ship hardened agents to production — Model Hardening Advisor plus HuggingFace dataset export
Hand a PR with proof — Review-ready workflow support and proof-ready lesson bundles a reviewer can verify in 30 seconds
Hand off without re-onboarding — Team lesson export/import for handoff or migration

What your Pro dashboard looks like

checks: 36 active
feedback: unlimited
exports: DPO + lessons

PAY-NOW PRO

Upgrade to Pro — $19/mo

Billed today · cancel anytime.

Team

$49/seat/mo

Shared enforcement memory for one team workflow first.

Everything in Pro for each seat
Shared lesson database for Team seats, with export/import and shared workflow rules
Org dashboard visibility across agent surfaces
Hosted review views for team verification evidence
Check template library for deploys, publish, refunds, and DB work
Workflow hardening sprint intake for approval boundaries and rollback safety
Email support during pilot rollout

Start 3-seat Team — $147/mo Or qualify first via Workflow Hardening Sprint intake →

Team is $49/seat/mo with a 3-seat minimum. Self-serve checkout starts you at $147/mo for 3 seats; add more seats from the dashboard. Prefer to qualify the workflow before paying? Start with the intake.

Need a custom diagnostic, sprint, or setup? Send workflow first →

Team Pilot

Start the AI Agent Governance Sprint with one repeat failure

This is the fastest path to first paid value for teams. Start with one repo, one workflow owner, and one blocker. The intake is designed to prove that ThumbGate reduces review churn, rollout risk, or repeated agent mistakes before a wider rollout.

Tell us the workflow. Get a proof plan.

The highest-fit Team buyer is already feeling one repeated failure. Start with a paid diagnostic when budget is ready, or use intake when you need qualification first.

30-minute intake

Workflow Hardening Diagnostic

Paid triage for one repeated agent or review failure, with a concrete proof plan and go/no-go recommendation.

$499

Pay for diagnostic

AI Agent Governance Sprint

48-hour Workflow Hardening Sprint for one workflow owner, one blocker, approval boundaries, rollback safety, one proof review, and one hardened rollout path.

$1500

Pay for sprint

Reliable AI Agent Governance Setup

Full ThumbGate-powered setup for Claude Code, Cursor, or Codex teams: npx install, custom prevention rules, audit trail, training, and proof reports. Optional ongoing rule management is $297/mo.

$3,997

Request setup

OpenClaw Agent Governance Kit

Self-serve digital kit with ThumbGate prevention rules, OpenClaw-compatible workflow prompts, agent registry checklist, proof report template, and Zernio launch copy.

$97

Buy kit

Not ready to pay from a checkout page? Send the workflow first. We can qualify the blocker, confirm the proof plan, and route you to the $499 diagnostic, $1500 sprint, or $3,997 governance setup only when the scope is real.

Send workflow first

FAQ

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.

How is ThumbGate different from model-training feedback loops?

ThumbGate's intelligence is context, not weights. It doesn't touch the model — 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 check blocks are hard enforcement, not soft suggestions.

What's the tech stack?

SQLite+FTS5 lesson DB for fast full-text search. MemAlign-inspired dual recall (principle-based rules + episodic context). Thompson Sampling for adaptive check 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 — no cloud required.

What AI agents and editors does this work with?

ThumbGate works with Claude Code, Cursor, Codex, Gemini CLI, Amp, Cline, OpenCode, and any other MCP-compatible agent. Cursor ships with a plugin bundle in this repo. Codex now ships both a standalone plugin bundle and a repo-local app plugin profile, and the published download is linked directly from this page. VS Code works when you run an MCP-compatible agent inside it, but this repo does not ship a standalone VS Code extension today.

Do I have to chat inside the ThumbGate GPT for enforcement?

No. The ThumbGate GPT is the ChatGPT entrypoint for checking proposed actions, capturing thumbs-up/down lessons, and getting setup help. Use it for advice and checkpointing; hard enforcement still runs locally where the agent executes after npx thumbgate init.

When should I use Pro versus the Workflow Hardening Sprint?

Start with the setup guide if you only need the local path. Choose Pro after one real blocked repeat when you want your own dashboard, DPO export, and proof-ready exports. Choose the Workflow Hardening Sprint when one workflow owner needs approval boundaries, rollback safety, and rollout proof before wider rollout.

How do we keep high-risk autonomous runs off the host?

ThumbGate is the control plane, not just a prompt layer. Workflow Sentinel predicts blast radius before execution, and risky local autonomy can be routed into Docker Sandboxes instead of running directly on the host. Team workflows also have a signed hosted sandbox lane for isolated dispatch when local repo access is not required.

How do we trust a new package release?

ThumbGate does not rely on vibes. Release-relevant PRs need a Changeset, SemVer rules keep version bumps honest, sync checks keep manifests aligned, proof lanes run before merge, and the exact main-branch merge commit is verified before the work is called done.

Do I need a cloud account?

No. Free keeps local enforcement on your machine with unlimited feedback captures, up to 5 active auto-promoted prevention rules, built-in safety checks, and hook blocking. Recall, lesson search, unlimited rules, and exports open up on Pro. No cloud account is required. The business starts when a team wants shared rules, approval boundaries, hosted review views, org dashboard visibility, and proof that survives handoffs. Pro is the optional solo side lane for a personal dashboard, DPO export, and team lesson export/import — share lessons across projects so one team's mistakes become every team's prevention rules.

What if my thumbs-down is vague?

For the current Claude auto-capture hook, ThumbGate can reuse up to 8 prior recorded entries and the failed tool call for a vague thumbs-down, then keep a linked 60-second feedback session open for later clarification. The timer resets when more context arrives, so the lesson stays attached to one feedback record instead of fragmenting into duplicates.

How are pre-action checks different from prompt rules?

Prompt rules are a starting point, not a finish line. Without prompt evaluation you do not know whether they still hold up under real tool use. ThumbGate adds the human-in-the-loop measurement loop and the enforcement layer: proof lanes, ThumbGate Bench, and self-heal checks show whether behavior improved, and Pre-Action Checks block the action before execution when it did not.

What does Pro cost?

Pro is $19/mo or $149/yr for individual operators and bills immediately through Stripe. Team is $49/seat/mo with a 3-seat minimum and starts through the workflow intake so scope, shared rules, and rollout proof are explicit before a team rollout.

Stop paying for the same AI mistake twice.

Block your first repeated AI mistake in 5 minutes.

See the footer before you ship the next repeat.

The fastest path for Claude users: install the extension and start blocking mistakes.

Use the GPT as a preflight desk for risky commands, refunds, deploys, and PR actions.

One gateway across the agent surfaces you already use

🧩 Claude Desktop Extension

⚡ Claude Code Skill

🤖 AI CLIs

☁️ Google Data Agent Kit

🧩 Codex plugin

🎯 Cursor plugin

✏️ Editor workflows

🗂️ MCP Server Directory

💬 ChatGPT GPT Actions

See the enforcement before you buy anything

🔍 Live Dashboard Demo

⛔ Check Reasoning Chains

📊 Org Dashboard (Team)

🧱 Isolated Execution Lanes

Data Processing Boundaries

🧪 Thompson Sampling

🪞 History-Aware Lessons