agentmail is a compliance and verification toolkit for AI agents, designed to fill the gaps in sanctions screening, KYA (Know-Your-Agent), and transaction fraud signals in autonomous agent payment scenarios. As agents begin to make autonomous payments through protocols such as x402, AP2, ACP, and Coinbase AgentKit, existing payment infrastructure typically assumes human oversight by default, and major payment platforms explicitly do not handle per-jurisdiction sanctions screening. agentmail provides agents with three capabilities they cannot accomplish on their own, exposed through three interfaces: MCP tools (for Claude Code / Cursor / Hermes), HTTP API (for any agent), and CLI (for developers), all three sharing the same core and data.
Main capabilities:
- Compliance layer: provides four tools that agents call before trusting or paying—sanctions screening, risk scoring, KYA verification, and dispute opening. The data covers 782 OFAC crypto wallets, 19,086 sanctioned names, and 16 embargoed jurisdictions. The data sources are public and free, require no keys, and the lists are cached locally in ~/.agentmail/cache/ with a 24-hour TTL, refreshed from the source; if the network is interrupted, the stale cache is used and degraded: true is reported via status(), avoiding silent failures. The osint provider performs exact and token-subset name matching, exact case-insensitive wallet matching, and ISO-2 country matching, with each match carrying a confidence level so users can decide how strongly to block.
- Verification toolkit: for email, it provides disposable inboxes that agents can use to register and receive OTP/magic link verification (powered by Mail.tm), with the flow being create_inbox() → address → fetch_code() → OTP; for SMS, it provides rentable phone numbers for phone/SMS verification, and the Mock provider can be used for development without a key, while setting AGENTMAIL_SMS_PROVIDER=fivesim and AGENTMAIL_FIVESIM_KEY=... allows connection to real numbers. Both share the otp.py extraction logic (regex matching of verification codes and magic links), outputting a unified {code, link} structure.
Usage: Users can use it through three approaches. MCP client approach (taking Claude Code / Cursor as an example): after configuration, the agent can call a total of 10 tools: sanctions_check, risk_score, kya_verify, dispute_open, create_inbox, fetch_code, create_number, fetch_sms, release_number, and list_inboxes. Note that the PyPI package name is sanctions-mcp (the agentmail name is already taken), but the Python import name is still agentmail. HTTP API approach: any agent can call it. CLI approach: for developers to use directly. Self-hosting is completely free and fully featured; you can run the MCP server locally or run the HTTP API on your own machine to screen against real OFAC data. The hosted API is located at https://agentmail-api.fly.dev，提供 API key authentication, rate limiting, and audit logs for each screening; the free tier is 50 checks per day with no registration required (by IP), and API keys are available for higher usage (Dev $19 per month, Team $99 per month, self-service checkout).
Use cases: It is suitable for scenarios where agents need sanctions screening, risk assessment, and identity verification before autonomous payments, transactions, or interactions with unknown counterparties, and also for development and testing workflows that require disposable email addresses or temporary phone numbers to complete registration and verification. The project also proposes the agentmail Sanctions Exposure Index (SEI), a proprietary framework with 5 factors for quantifying AI agents' OFAC sanctions exposure (scores from 10 to 1000), with the full report and interactive calculator at sanctionsai.dev, licensed under CC BY 4.0.
Limitations and caveats: Self-hosting is a free tier rather than a trap; what is paid for is uptime, data freshness, and audit logs, while the data itself remains public. 5sim numbers may be shared after release, making them suitable for receiving one-time OTPs, but absolutely unsuitable for two-factor authentication on accounts intended to be kept long term. The project uses the MIT license. Other features, pricing details, or platform support not mentioned here are not yet provided in the materials.
