The full upstream README, mirrored here for reference. Install config, tool schemas, adoption signals, and an original overview live on the Agent Commerce Guard listing page.
Approval gates and receipts for AI agents that can spend money, sell services, post socially, launch tokens, deploy publicly, or touch wallets.
Agent Commerce Guard is a local-first CLI and skill pack for builders adding x402, ACP, wallets, seller runtimes, paid APIs, or social integrations to autonomous agents.

Try the free evaluator | Preview the paid package | Add the GitHub Action | Buy the private package for 1 USDC | Get support
The live evaluator includes one-click presets for wallet spend, production deploys, token launches, and marketplace jobs, so each decision path can be tested without writing a manifest first.
| Free hosted preflight | Private local package | |
|---|---|---|
| Best for | Public CI action manifests | Private or sensitive agent workflows |
| Execution | Hosted evaluator, up to 5 actions / 20 KB | Local CLI, HTTP evaluator, skills, policies, and templates |
| Price | Free | 1 USDC on Base |
The checkout links to the full sample report, Base MCP purchase recipe, and support before payment, then verifies the Base USDC transaction server-side. Both the checkout and success handoff retry pending receipts automatically before the private package unlocks. A submitted transaction hash is preserved in the current URL so a refresh resumes verification instead of losing the unlock path.
The public package proof shows the exact CLI report format and file inventory before payment. Canonical tests compare the samples to the current policy engine and verify each advertised file against the paid tarball:
Install a non-executing preflight workflow for public, non-sensitive action manifests:
The skill enforces the hosted-data boundary, evaluates at most five actions and 20 KB through the free endpoint, and hands private workflows to the paid local package without initiating a payment.
Install the policy-only review plugin directly from GitHub:
The plugin's skill uses /api/policy-review, returns only decisions and
required evidence, and includes no checkout or payment links. Its immutable
1.0.0 source is tagged
copilot-plugin-v1.0.0.
The
official Awesome Copilot review
is ready for review after passing its automated quality gates.
Connect an MCP client to the public, stateless Streamable HTTP endpoint:
Codex:
Claude Code:
VS Code mcp.json:
Setup and verification: Vercel guide or GitHub Pages mirror.
The server exposes one read-only, non-destructive tool,
review_agent_commerce_actions. It accepts up to five non-sensitive public
actions and 20 KB, returns decisions and required evidence, and does not
execute, pay, sign, deploy, post, or include purchase instructions in tool
output. MCP 1.2.0 also exposes one public JSON resource,
https://agent-commerce-guard.vercel.app/offer.json, so clients can discover
the exact 1 USDC Base offer without mixing purchase data into policy-tool
results. Server initialization points clients to that resource, requires
explicit approval before payment, and requires Base transaction verification
after payment. The review-before-purchase prompt joins policy review, offer
inspection, explicit approval, and receipt verification without initiating a
payment. Registry metadata is available in server.json.
The official Registry lists
io.github.fxjim/agent-commerce-guard
as an active remote server.
Install the free public buyer workflow for Codex, Claude Code, Cursor, and other Agent Skills-compatible assistants:
The same skill is also available from the canonical product domain:
The skill preflights the public offer and requires explicit approval before any
Base Account, Base MCP, x402, browser-wallet, or manual payment. Its source is available at
skills/buy-agent-commerce-guard/SKILL.md.
Its indexed directory page is available on
skills.sh.
The paid product package remains available only after a verified 1 USDC
payment.
Run the free hosted policy preflight in an ubuntu-latest workflow:
The manifest may be a single action, an action array, or an object with an
actions array. The Action sends at most five actions and 20 KB to the public
hosted evaluator, prints sanitized decisions, and can fail on denied actions or
actions requiring review. It rejects obvious secret-bearing JSON keys before
the request.
The Action exposes checkout-url for the private local package,
x402-evaluate-url for the paid workflow API, sample-report-url for the full
CLI example, package-metadata-url for the machine-readable inventory, and
base-mcp-recipe-url for an authorized Base MCP assistant. The job summary
links to the proof and purchase paths after every completed evaluation.
Do not submit secrets, private source, credentials, or sensitive customer data.
Use the paid local package when an action manifest cannot leave the runner.
See examples/github-actions/ for a complete
workflow and harmless example manifest.
The launch pass is available for 1 USDC on Base.
/pay page leads with a user-confirmed 1 USDC Base Account payment, automatically waits for a pending receipt, preserves the returned transaction hash across refreshes, and verifies it server-side before unlock.c67fb09bd83da591c58ae5fae002a6a59557fc97/verify?tx={tx} and /success?tx={tx} as post-payment unlock pages.llms.txt, agents.json, agent.json, offer.json, .well-known/agent-commerce-guard.json, .well-known/offer.json, openapi.json, .well-known/openapi.json, .well-known/ai-plugin.json, package-metadata.json, .well-known/package-metadata.json, status.json, .well-known/status.json, .well-known/x402, .well-known/agents.json, .well-known/agent.json, .well-known/purchase.json, pay/, robots.txt, and sitemap.xml.agent-commerce, agent-safety, guardrails, approval-gates, codex, x402.Agent commerce is moving faster than agent governance. This tool blocks or escalates actions that cross financial, production, social, token, or credential boundaries until there is explicit evidence and approval.
https://agent-commerce-guard.vercel.app/offer.json.https://agent-commerce-guard.vercel.app/pay and use the official Base Account control to review and confirm 1 USDC on Base mainnet.https://agent-commerce-guard.vercel.app/base-mcp.json in an authorized Base MCP assistant, read https://agent-commerce-guard.vercel.app/pay.json for the exact manual Base payment request, or scan https://agent-commerce-guard.vercel.app/qr./verify?tx={tx}, /success?tx={tx}, or /api/base-payment-status?tx={tx}. Both browser pages automatically retry a pending Base receipt and resume from the URL after a refresh./success?tx={tx} or /api/base-download?tx={tx}.Agent buyers can also call https://agent-commerce-guard.vercel.app/api/x402-download directly and satisfy the returned x402 payment requirement.
Agent workflows can call POST https://agent-commerce-guard.vercel.app/api/x402-evaluate and satisfy the returned x402 payment requirement to receive a sanitized policy decision.
The x402 endpoints return Base-mainnet exact payment requirements for 1000000 atomic USDC and include route-level Bazaar metadata for agent discovery.
Purchase questions, wallet-flow issues, and x402 integration notes can go in the pinned launch/support issue: https://github.com/fxjim/agent-commerce-guard/issues/1.
Launch context and buyer discovery thread: https://github.com/fxjim/agent-commerce-guard/discussions/2.