Click any agent command to evaluate it against real ThumbGate policies in < 5ms:
Wire local hooks:npx thumbgate quick-start
π First mistake blocked! Run npx thumbgate quick-start locally or upgrade to Pro ($19/mo).
Finish setup
Finish setup to record your first gate block β
Run npx thumbgate quickstart to fire your first PreToolUse block and unlock the built-in safety rule pack + 3 custom block rules on the Free tier (7-day local Pro trial included).
What's happening right now? Search memories, inspect active checks, manage your team, and export training data. This is your control plane for AI agent behavior.
π Search β find any memoryπ‘οΈ Gates β what's blockingπ₯ Team β org metricsπ€ Export β DPO training data
π Demo Mode β sample data. Pro unlocks your personal dashboard with search, DPO export, and gate analytics.Start Pro now
No human feedback recorded on this hosted account yet.
Local feedback history is not automatically synced here. On the machine where you use ThumbGate, run npx thumbgate dashboard --open to view your local feedback. Configured gates are available rules; lessons are independent records, not human feedback.
π¬ Chat with your data
Ask about your gates, blocks, feedback, and lessons β answered locally from your own data, no cloud. Each AI reply shows its exact model & privacy route.
Powered by local dashboard data. Optional Gemini/Perplexity keys only expand open-ended analysis.
π Since Last Review ?Shows new feedback, negative signals, lessons, and blocks registered since your last baseline checkpoint.
No review checkpoint yet. Mark the current dashboard as reviewed to start seeing only new changes.
π‘οΈ How Pre-Action Gating Works (Interactive Flow Diagram)
Recorded blocked entries (latest 20 in 30 days)
Loading gates...
Shared Team Reliability
See which agents drift, which gates save the team the most time, and whether the shared workflow is getting safer.
Loading team metrics...
Loading predictive insights...
Highest-Risk Agents
Loading agent risk...
Top Blocked Gates
Loading blocked gates...
Predictive Watchlist
Loading predictive watchlist...
Agent Surface Inventory
See which tools, policy sources, and MCP surfaces are actually active before you widen rollout.
Loading inventory...
Observed Tools
Loading observed tools...
Policy Sources
Loading policy sources...
Background Agent Mode
Queued and scheduled agents need explicit checkpoints before they touch code, money, or customer systems.
Loading background-agent mode...
Governance Pressure
Loading background-agent governance...
Run Types
Loading background-agent run types...
Regulated Buyer Proof
Show where policy came from, how many runs were reviewed, and which proof artifacts are ready when buyers ask for auditability.
Loading proof posture...
Policy Origin Proof
Loading policy origin proof...
Latest Replay Artifacts
Loading replay artifacts...
Governed Data Chat
Ask questions over your local ThumbGate data (lessons, raw feedback via LanceDB vectors, gate stats, receipts). Uses local retrieval plus your configured LLM, including open-source/local endpoints through THUMBGATE_LOCAL_LLM_ENDPOINT. Dialogflow is not the chatbot layer; DFCX/Vertex cards are optional deployment-readiness checks for buyers who run their own Google agent stack.
Chat With Local ThumbGate Data
Connect your dashboard, then ask about local ThumbGate data.
Runtime Readiness
Loading enterprise status...
The chat itself is local RAG plus your LLM. Open-source/local models are the default enterprise path; DFCX/Vertex readiness is optional evidence for teams that already operate Google agents in their own tenancy.
AI Component Inventory
This API scans the dashboard serverβs working directory. On the hosted dashboard it describes the server checkout, not projects on your machine. Run the CLI export below from your project for its inventory.
Enterprise evidence for AI/ML usage: provider SDKs, agent frameworks, vector databases, Vertex/Gemini/Dialogflow CX references, and local model artifacts. Export this as ML-BOM proof before claiming an AI system is production-ready.
A constrained JSON spec builds opinionated review dashboards from approved cards, lists, and callouts.
Loading generated views...
Loading generated view...
Policy Origins
These are the dashboard serverβs resolved settings. To inspect your projectβs local setup, run npx thumbgate doctor in that project.
Loading settings hierarchy...
Loading settings summary...
Layer Precedence
Loading layers...
Routing Preview
Loading routing...
Resolved Setting Origins
Loading settings origins...
Curated Gate Templates
Loading template library...
Loading templates...
π Feedback Insights & Lesson Pipeline
How your thumbs-up/down signals turn into lessons that prevent repeat mistakes.
πΈ Estimated tokens saved
Computed from your real blocked-action count Γ 2,000 input + 600 output tokens per avoided retry, priced at a Sonnet-heavy blend (80% Sonnet 4.5 / 15% Opus 4.6 / 5% Haiku 4.5). Conservative estimate β actual savings may be higher.
β
β tokens Β· from β blocked calls
Feedback β Lesson Pipeline
Loading pipeline...
Lessons Generated (30 days)
Gate Effectiveness (14-day audit)
Highest-ROI Next Actions
These recommendations come from real feedback, risk patterns, and delegation pressure in your current runtime.
Loading next actions...
How ThumbGate Learns
ππ
1. You React
Thumbs-down on a bad action, thumbs-up on a good one
π§
2. Lesson Distilled
ThumbGate extracts what went wrong and how to avoid it
π‘οΈ
3. Gate Promoted
Repeated lessons auto-promote into pre-action checks
π«
4. Mistake Blocked
Gates intercept the same mistake before it happens again
Export Training Data (DPO)
What is DPO?
Direct Preference Optimization is a technique for fine-tuning LLMs using human preference data. Your π/π feedback is converted into training pairs:
π "chosen" β the response that worked
π "rejected" β the response that failed
Use these pairs to fine-tune any model (OpenAI, Llama, Mistral) so it actually learns from your corrections β not just flags mistakes at the gate, but stops making them.