A thoughtfully engineered local-first gating layer: no LLM on the enforcement path, auditable decisions, break-glass recovery, and honest disclosure of limits. Its value hinges on hook interception — hard enforcement only on PreToolUse-capable agents — and the docs candidly note advisory-mode limits, estimate-based savings, and unshipped roadmap items. Well suited to teams wanting a cross-agent, learning rule store; overkill for simple static denylists.
ThumbGate is a local-first Pre-Action Checks engine that runs as a stdio MCP server and via PreToolUse hooks to evaluate proposed AI agent tool calls before execution. It captures thumbs-up/down feedback as local lessons, promotes repeated failures into prevention rules, and hard-blocks detected secret leaks plus direct self-disable commands (process kill, env bypass) by default. Other high-risk classes (rm -rf, force-push, fetch-and-run) warn and log by default; THUMBGATE_STRICT_ENFORCEMENT=1 preserves deny decisions for every matched blocking rule. Gate decisions are deterministic (literal/AST matching with local bge-small embeddings via LanceDB) — no LLM runs on the enforcement path. Claude Code, Codex, and Gemini CLI get hard interception; Cursor, Cline, OpenCode, and Amp get advisory gate_check only. Free tier: 2 feedback captures/day (10 total), 3 active auto-promoted prevention rules; Pro at $19/mo adds unlimited rules, lesson recall/search, and a personal dashboard. MIT-licensed.
Run npx thumbgate init to auto-detect and wire your agent's integration, or run it as a stdio MCP server directly: npx -y thumbgate serve. Optional THUMBGATE_MCP_PROFILE (essential | default | readonly | locked) overrides the tool profile; THUMBGATE_STRICT_ENFORCEMENT=1 enables strict blocking. Do not use npm start for MCP — that launches the HTTP API, not the stdio server.
{"mcpServers":{"thumbgate":{"command":"npx","args":["-y","thumbgate","serve"]}}}