Base AgentGuard

Base AgentGuard is an open-source safety studio for AI agents that use Base MCP. It helps builders review wallet actions before approval, score risky agent intents, generate send_calls-ready receipts, and document custom plugin flows for Base mainnet and Base Sepolia.

The project follows the Base AI agents documentation pattern:

• connect an assistant to Base MCP
• detect the user's Base Account wallet
• prepare unsigned calldata or payment intent
• map the prepared response to send_calls
• require user approval for every write action

Official Base references:

• Base MCP overview: https://docs.base.org/ai-agents
• Base MCP quickstart: https://docs.base.org/ai-agents/quickstart
• Custom plugins: https://docs.base.org/ai-agents/plugins/custom-plugins

Why this project exists

AI agents can now help users send funds, swap tokens, sign messages, execute contract calls, and pay x402-enabled APIs. Those workflows are powerful, but they need a visible review layer. Base AgentGuard is that layer for builders: it turns an agent request into a readable intent receipt before the user approves anything.

Features

• Base mainnet and Base Sepolia RPC health checks
• Agent action scoring for sends, swaps, contract calls, signatures, and x402
• x402 payment caps and endpoint review receipts
• Builder Code attribution planning for agent transactions
• High-risk term detection for approvals, permits, upgrades, bridges, claims, proxies, and admin calls
• Base MCP receipt preview for send_calls
• Custom plugin markdown spec in plugins/base-agentguard.md
• Solidity registry sketch in contracts/BaseAgentIntentRegistry.sol
• OpenAI OSS program application copy in docs/openai-oss-application.md
• Base launch and funding playbook in docs/base-launch-playbook.md
• Impact tracking in docs/impact-metrics.md
• No API keys, no custody, no seed phrase flow, and no hidden transaction execution

Quickstart

npm start

Then open:

http://localhost:4173

For a no-server preview, open index.html directly in a browser.

Validate

npm run validate

The validation script checks that the static app, plugin spec, example receipt, license, and docs are present and internally consistent.

Repository layout

.
|-- index.html
|-- styles.css
|-- app.js
|-- data/risk-rules.json
|-- examples/send-calls.receipt.json
|-- plugins/base-agentguard.md
|-- contracts/BaseAgentIntentRegistry.sol
|-- docs/openai-oss-application.md
|-- docs/openai-maintainer-evidence.md
|-- docs/base-build-notes.md
|-- docs/base-launch-playbook.md
|-- docs/impact-metrics.md
|-- docs/security-and-trust.md
|-- docs/source-review.md
|-- ROADMAP.md
|-- CONTRIBUTING.md
|-- SECURITY.md
|-- LICENSE
`-- tools/

Suggested Base builder path

1. Publish this as a public GitHub repository.
2. Deploy the static app to Netlify, Vercel, or GitHub Pages.
3. Deploy BaseAgentIntentRegistry.sol to Base Sepolia.
4. Add the deployed contract address and transaction hash to this README.
5. Register a Builder Code on Base.dev and add it to the attribution plan.
6. Create issues for risk rules, x402 review, plugin examples, and Base Account onboarding.
7. Share the project as a Base AI agent safety tool, not as an airdrop claim page.

Security posture

This MVP is read-only by default. It does not custody funds, does not ask for private keys, does not execute swaps, does not sign messages, and does not send transactions. It produces review receipts and examples that a user or assistant can inspect before moving into Base MCP approval flows.

License

MIT