A pre-trade risk API is only as useful as it is reachable. For a while, AgentRisk M2M was a solid, well-tested HTTP endpoint — but reachable only if a developer already knew it existed and was willing to write the integration code by hand. That's a real barrier, and today it's gone.
The gap
Every major framework for building autonomous agents ships its own convention for adding a capability: an MCP tool, an ElizaOS action, a LangChain tool, a Vercel AI SDK tool, a Coinbase AgentKit action provider. A developer working in any of these doesn't want to hand-roll an HTTP client and parse a JSON response — they want npm install and a working tool in their agent's toolset within a minute.
What shipped
Five packages, each a thin, tested wrapper around the same underlying API (agentrisk.dev/scan) — no duplicated logic, no drift between them:
agentrisk-mcp — a standard MCP server, works with Claude Desktop, Cursor, and any MCP-compatible client. npm install -g agentrisk-mcp, one config block, done.
@agentrisk/plugin-elizaos — an ElizaOS action (CHECK_TOKEN_RISK) that fires when a message mentions a Base address alongside words like "safe," "honeypot," or "rug." Drops straight into a character file's plugin list.
@agentrisk/langchain — a DynamicStructuredTool for LangChain agents, addable to any tool list like any other LangChain tool.
@agentrisk/ai-sdk — a Vercel AI SDK tool using the current inputSchema / execute shape (the API changed between major versions of the ai package — worth checking your installed version before copying examples from older docs).
@agentrisk/agentkit — a Coinbase AgentKit action provider, the one I'd call the most consequential integration here, since AgentKit is Coinbase's own first-party framework and sits in the same ecosystem as the x402 facilitator and Builder Codes I'm already using for payments and attribution.
What each one actually does
Every package calls the same real-time analysis: honeypot detection, deployer wallet freshness, brand impersonation checks, on-chain LP-lock verification, and a live sell simulation across every DEX architecture live on Base — Uniswap V2, V3, V4, and Aerodrome including Slipstream concentrated liquidity. That last part matters more than it sounds: a token can pass every static check and still be unsellable the moment real liquidity conditions are queried, which is exactly the gap a code-only scanner can't see.
A genuinely annoying bug along the way
Building the AgentKit package surfaced a real dependency conflict worth flagging for anyone doing the same: @coinbase/agentkit pins its internal type expectations to Zod 3, while a fresh npm install zod today pulls Zod 4. The compiler error (missing properties from type 'ZodType'... _type, _parse, _getType) gives no hint that version mismatch is the cause. Pinning zod@3.25.76 explicitly fixed it immediately.
Pricing, unchanged
0.15 USDC per scan via x402, no API key, no subscription. Free public checks at agentrisk.dev for anyone who wants to see the output shape before wiring up a package.
Where it stands
Five repos, five packages, one API behind all of them. Whichever framework an agent is built on, the integration path is now identical to installing any other tool in that ecosystem — no special-casing, no manual HTTP client required.
Repos: github.com/Neurobyteio/agentrisk-mcp, agentrisk-eliza, agentrisk-langchain, agentrisk-ai-sdk, agentrisk-agentkit
API: github.com/Neurobyteio/agentrisk
Try it: agentrisk.dev
Top comments (0)