arcjet

Create one Arcjet client at module scope with Shield as a base rule, then call protect() inside each route handler with a rate limit rule like tokenBucket. Use characteristics like userId to key limits per authenticated user rather than by IP.

Use Arcjet Guard instead of request-based protection, since tool calls have no HTTP request. Call guard() with a hardcoded label and per-tool rate limit rules, or use the official adapter for frameworks like Vercel AI SDK, Mastra, or LangChain.

Use Guard for MCP servers because MCP tools are invoked over stdio or SSE, not HTTP requests. Request-based protection with protect() only applies when an actual HTTP request object exists, such as in Express or Next.js route handlers.

Arcjet supports JavaScript/TypeScript, Python, and Go server-side code. JS frameworks use packages like @arcjet/next and @arcjet/guard, Python uses the arcjet package, and Go uses github.com/arcjet/arcjet-go. Other languages are not supported.

JS and Go HTTP rules default to DRY_RUN mode, so you must pass mode: "LIVE" or ModeLive to enforce. Also verify a real request triggered a decision by checking the CLI with arcjet requests list or the Console, since a missing decision is not a denial.

The default WASM backend detects only email, phone, IP, and credit card numbers. For names, addresses, and government identifiers, add the on-device Rampart NER backend via @arcjet/sensitive-info-rampart or the Python equivalent, and pass it as the rule's backend.