feat(skills): gate-my-tool — one-command install of a CredentAgent gate

[TheBlackBit]

In plain terms

This adds a recipe file that a coding agent (like Claude Code) can follow to protect an online shop's "buy" button in one run. A developer with an existing agent-facing server can say "gate my checkout tool", and the agent wires in CredentAgent so the purchase only completes after the human proves the required credential (their age, a payment authorization) from their phone — instead of the developer reading the docs and wiring it by hand.

What you're approving

• One new file: .claude/skills/gate-my-tool/SKILL.md — a "skill", i.e. a packaged instruction set coding agents in this repo can invoke. No library code changes, nothing ships to npm, no user-facing change.
• Worst case if the instructions were wrong: an agent following them would wire the gate incorrectly. That risk is covered — the skill was verified by following it step by step against a sample server (see below), and it requires the security-bypass test as part of the install.
• Closes Claude skill gate-my-tool — one-command install of a CredentAgent gate #17 (the roadmap item asking for exactly this one-command install).

How to test

Read the skill and check the code snippets against the real API:

cat .claude/skills/gate-my-tool/SKILL.md

Then confirm the repo still builds and tests green (the change is documentation-only):

npm run build && npm run test

Expected: build succeeds; 38 test files, 605 tests pass.

---

For reviewers — the detail

What the skill teaches, step by step. Pointed at an MCP server (a server exposing tools to AI agents) whose registerTool handler completes a consequential action ungated:

1. Locate the settling write (the fulfillment/charge) — that is what moves behind the gate.
2. Wire credentagent.orders.serve(app) once at startup, and move fulfillment into the order.settled event.
3. Rewrite the tool handler to price the order server-side from the host's catalog (security invariant 2 — never trust client input), open a gated order with credentagent.orders.create({ order, policy }), and return the approve link + requires manifest instead of completing.
4. Add a get-order-status poll tool over credentagent.orders.retrieve (the three execution contexts: tool mints the link, the human proves on the page, the agent polls).
5. Write the bypass test (delegating to the existing write-bypass-test skill): an unverified age-restricted checkout must stay pending and unfulfilled — and the test must go red if the gate is reverted.

It also carries the honesty fence (trust_level: "presence-only-demo" — a flow demo, not a real safety control) and a red-flags table mirroring the style of the existing add-ceremony-rail skill.

Decision table: the skill defaults to the orders facade and points hosts that already own their catalog/order stores and checkout route at defineHost(...) + examples/bring-your-own-host.mjs.

How it was verified (the issue's "done when", executed). A fixture MCP server with an ungated checkout was built; baseline confirmed it settled a 21+ item with zero proof. Applying the skill's steps verbatim produced a gated tool where the same unverified checkout returns the approve URL + manifest, retrieve() stays { ok: false, pending: true }, and fulfillment never runs. Load-bearing proof: reverting the handler to the ungated write turned the bypass test red; restoring it, green. Every API call in the snippets was checked against the source (orders.create → { id, approveUrl, manifest }, empty-id minting in orders.ts, the GateOrder shape in types.ts) and the MCP result shape against SDK 1.29.

Also in this change: issue #17 itself was rewritten (title + body) for the CredentAgent rename and to drop its stale claim that the old discovery surface (llms.txt + /.well-known/attestomcp.json) still ships — it was removed during the rename.

🤖 Generated with Claude Code

[vercel]

The latest updates on your projects. Learn more about Vercel for GitHub.

Project	Deployment	Actions	Updated (UTC)
credentagent-demo	Ready	Preview	Aug 7, 2026 3:05am