Skip to content

v0.1.0-beta — public beta

Pre-release
Pre-release

Choose a tag to compare

@agent443-dot agent443-dot released this 10 Nov 23:53
· 2 commits to main since this release
d6aee65

The first public beta. AgentGuard runs with real money on Robinhood Chain (USDG) and Solana (USDC) mainnet. It is self-hosted: each user runs their own guard-core, Telegram bot and signing page, and their vault is owned by their own Phantom wallet. It has had internal reviews but no external audit yet. Hold in the vault only what you can afford to lose, and keep the daily cap small.

Install: Install the beta in the README, or docs/self-host.md by hand. How it works: the technical whitepaper.

What is in it

  • Onchain cap, no custom contract. A Squads v4 spending limit on Solana and the Safe Allowance Module on Robinhood Chain. The guard's hot wallet can draw at most the daily cap.
  • Every payment checked. guard-core holds the policy for each agent:
    • per-payment and daily limits;
    • token, chain, recipient and domain allowlists;
    • a simulation of the exact transfer;
    • your approval in Telegram above a threshold.
  • Pinned x402 payees. A payment to a new recipient for a known site, or at more than twice the last price, is flagged and waits for you.
  • Circuit breaker. An agent that makes too many payments in a minute is paused, and you are told.
  • Shadow mode. New agents start in shadow mode: they run the whole check and pay nothing, so you see what they would have spent. /mode <agent> live turns one on.
  • Task budgets. An agent can ask for one budget for a whole task: an amount, sites, a time limit. It never raises the policy.
  • Kill switch. /kill in Telegram stops guard-core at once. One Phantom signature per chain then removes the onchain cap.
  • Proof of cap. A read-only page shows anyone the most an agent can draw in the next 24 hours, read from the chain.
  • Owner commands in Telegram: /status, /budget, /report with a weekly report, /doctor, /agent, /policy, /allow, /deny, /mode, /kill, /resume.
  • Setup wizard: create-agentguard, with --dry-run and doctor. It also runs on any server with docker-compose.yml.
  • MCP tools for agents: paid_fetch, transfer, check_intent, get_budget, start_task, and a setup command for Claude Code, Claude Desktop, Cursor and others.

Built in, off until you configure them

  • Signed receipts: an Ed25519 receipt for each payment, with RECEIPT_SIGNING_KEY_FILE.
  • Passkey approvals: payments above PASSKEY_ABOVE can be approved only with your passkey, so a stolen Telegram account can reject but never approve them.

Known limits

  • No external audit yet. The internal reviews are in docs/security-review.md, and the known risks in docs/final-report.md.
  • The owner key is in a software wallet (Phantom). The signing page is a trust root: deploy your own, from your own protected main.
  • Phantom may call the Solana multisig creation "malicious" on a new domain. See docs/troubleshooting.md for what to check before you confirm.
  • The packages are not on npm yet. Run everything from a checkout of this repository.
  • Base is next. It runs on Base Sepolia for development only.