| What it intercepts |
MCP requests flowing through Cloudflare's portal, evaluated against tool-specific policies before passing or blocking |
Tool calls the agent harness exposes (bash, SQL, file write, HTTP fetch, MCP tool): mandatory PreToolUse hooks where supported; MCP gate_check / advisory rules on MCP-only adapters |
| Where the gate runs |
Cloudflare's network — a centralized portal layer in front of MCP servers |
Locally, inside the developer's agent session; no cloud dependency on the enforcement path |
| Action classes covered |
MCP operations only — anything that does not flow through the portal is out of scope (local shell, file system, git, direct database clients) |
All tool-call classes the agent harness exposes, MCP included |
| Risk taxonomy |
read-only (no risk) → contained write (e.g. creating merge requests) → critical (production deployments, bulk deletions) |
👍 allow → 👎 warn + log → ⛔ hard-block; strict mode turns warns into denies. Maps 1:1 onto WriteGuard's tiers |
| Rule source |
Tool-specific policies authored centrally — the stated goal is implementing them once at the portal instead of re-implementing per MCP server |
Operator feedback (👍/👎) distilled into lessons and auto-promoted to prevention rules only after passing precision/recall gates; each rule traceable to the lesson that produced it |
| Identity model |
Existing OAuth credentials, with agent context added to downstream audit logs — no standalone agent accounts |
Per-agent inventory and per-gate hit rates on /dashboard; agent id and metadata captured on every gate decision |
| Audit output |
Portal audit trail with agent labels, omitting values for keys considered secret or sensitive |
Local audit log with rule version, source lesson, decision, and timestamp; signed execution receipts per attempt (see the receipt schema) |
| Availability |
Private beta; no public cost information disclosed |
MIT-licensed npm package, installable today: npx thumbgate init; hosted plans on /pricing |