Integrate Arcjet Guard Genkit

Integrate Arcjet Guard into a Genkit app

@arcjet/guard's Genkit v1 namespace wraps the agent's existing Arcjet client. It never talks to the Arcjet API itself. Three surfaces, one decision rule:

• An authored tool (ai.defineTool(config, handler)) → guardTool(). After defineTool the object is a ToolAction; generate() calls it as a function. DENY returns a structured ArcjetDenialResult. Do not throw. Do not call interrupt(). Do not throw ToolInterruptError.
• Filesystem / MCP / unwrapped tools → guardMiddleware(). A generate({ use }) middleware whose tool hook is the generate()-wide gate. It denies by returning a completed ToolResponsePart without calling next(). Already-branded tools are skipped when they can be looked up. Requires the generateMiddleware tool hook (Genkit >= 1.33).
• Correlation → genkitContext() reads a field the integrator put on generate({ context }) / the tool handler's { context } (correlationId, then sessionId, then conversationId, then a caller-owned flow / run id). It never mints a new id. It never reads traceId. It never treats interrupt / resumed as correlation.

This namespace is JS genkit() + ai.defineTool + ai.generate. Not Go / Python Genkit. Do not also wrap the same tool with @arcjet/guard/vercel-ai/v7 or @arcjet/guard/claude-managed-agents/v0. Zod is Genkit's, not ours.

Screen user text before generate() — there is no inbound hook. Middleware model is not Guard.

There is no first-class inbound channel, so there is no guardInbound. Put prompt-injection (and other inbound rules) in the application before ai.generate() / chat.send(). The middleware model hook intercepts the model call, not user text. It is not this policy gate.

interrupt() / defineInterrupt / toolApproval are HITL, not a policy gate.

interrupt() / defineInterrupt / @genkit-ai/middleware toolApproval / restartTool / finishReason === "interrupted" is human-in-the-loop. Same trap as Mastra requireApproval, Claude canUseTool, LangGraph interrupt(), OpenAI Agents needsApproval, and Google ADK requireConfirmation. There is no guardApproval. Do not wrap them as Guard.

Deny inside defineTool (and guardMiddleware's tool hook). MCP and filesystem-injected tools skip an unwrapped handler.

The authored defineTool handler is the deny point for tools you own. Filesystem middleware tools, MCP tools, and anything not wrapped with guardTool skip that handler. guardMiddleware is the generate()-wide gate for those. returnToolRequests: true means the app calls the tool itself — guardTool on the defineTool handler still gates that; guardMiddleware does not run if they never generate() the tool.

guardMiddleware can deny. Genkit's resolveToolRequest treats a ToolResponsePart returned without calling next() as a completed tool result. Throwing ToolInterruptError sets finishReason: "interrupted" (HITL — do not do this). Returning undefined drops the tool request (do not do this).

generate({ use }) must receive a plain object { name, instantiate }. A raw function becomes a model hook only. A function with instantiate

• plugin throws “must be called with ()”.

Questions to ask the human first

Ask only what you cannot infer from the code; suggest defaults.

1. Which tools are risky (external side effects, irreversible, spends money, sends messages)? Those get guardTool. MCP / filesystem / tools you did not author get guardMiddleware.
2. What limits? (e.g. "10 lookups/min per order" → tokenBucket.)
3. Who is the user for metadata — an opaque user/tenant ID (never PII)? Default: none. Pass it via metadata on the policy. Put the conversation / session id you already have on ai.generate({ context: { sessionId } }) and on guardMiddleware({ sessionId }) — the tool hook does not receive ALS context today. That id is the correlation id, not the user.
4. Is an Arcjet outage unacceptable? Every helper defaults to onGuardError: "deny". Ask explicitly about inbound screening before generate(): failing closed there means the agent does not run for the duration of the outage, so "allow" is a routine and legitimate choice at that one call site.

The six things readers get wrong

1. There is no guardInbound. Screen prompt injection before ai.generate() / chat.send(). Middleware model is not Guard.
2. interrupt() is not a policy gate. It is HITL. Use guardTool or guardMiddleware. A denial is a completed toolResponse, not finishReason: "interrupted".
3. The import path is versioned and there is no alias. @arcjet/guard/genkit/v1. @arcjet/guard/genkit does not resolve.
4. Correlation is read, never minted. Do not call createAgentContext inside a generate / tool callback — that generates a second id and splits the Sequence. Put the id you already chose on generate({ context }). Do not read Session.sessionId from a Session constructed without an id — that class mints a UUID. Do not use traceId (OTel / Genkit mints one). Do not treat interrupt / resumed as correlation.
5. Do not double-wrap with @arcjet/guard/vercel-ai/v7. guardTool throws if the tool already carries the Arcjet protection brand.
6. A denial from guardTool is a structured object, not a throw. Wrap the returned ToolAction (the callable generate() invokes), not the inner handler. outputSchema validation runs inside action(). Wrapping outside means DENY returns ArcjetDenialResult without schema check, so the model still sees a completed tool result. Wrapping the inner handler would throw on schema mismatch and fail generate().

Step 1: Install and find the guard client

Install @arcjet/guard (required), plus genkit (optional peer, needed for @arcjet/guard/genkit/v1). Always use the versioned path: @arcjet/guard/genkit/v1 resolves; @arcjet/guard/genkit throws ERR_PACKAGE_PATH_NOT_EXPORTED. Zod is Genkit's peer, not ours — install zod only if the app already uses it outside genkit. Public API is import { genkit, z } from "genkit". Node 22+.

npm install @arcjet/guard genkit

If the agent has no guard client yet, launch one once at module scope:

import { launchArcjet } from "@arcjet/guard";

export const arcjet = launchArcjet({ key: process.env.ARCJET_KEY! });

Step 2: Gate authored tools

import { genkit, z } from "genkit";
import { guardTool } from "@arcjet/guard/genkit/v1";
import { tokenBucket, localDetectSensitiveInfo, policyInput } from "@arcjet/guard";

import { arcjet } from "./arcjet.js";

const ai = genkit({/* plugins, default model */});

const lookupLimit = tokenBucket({
  refillRate: 10,
  intervalSeconds: 60,
  maxTokens: 10,
});
// Factory then text — same shape as `detectPromptInjection()(text)`.
// Scan free-text args (a note, reason, body). An opaque `orderNumber`
// will not trip EMAIL / phone / card / IP, so do not pass it here.
const detectPii = localDetectSensitiveInfo();

export const lookupOrder = guardTool(
  arcjet,
  ai.defineTool(
    {
      name: "lookup_order",
      description: "Look up an order by number",
      inputSchema: z.object({
        orderNumber: z.string(),
        note: z.string(),
      }),
    },
    async ({ orderNumber, note }) => ({ orderNumber, note, status: "shipped" }),
  ),
  {
    action: "order.looked-up",
    actor: userId,
    inputs: (input) => ({
      orderNumber: policyInput.server.string(input.orderNumber),
    }),
    rules: (input) => [
      lookupLimit({ key: input.orderNumber, requested: 1 }),
      detectPii(input.note),
    ],
  },
);

• Omit rules to submit none. The guard call still happens.
• Optional actor and inputs (static, or a resolver over this adapter's native call — parsed input plus trusted runtime/context) are forwarded on the guard call so a remote policy that declares those names can evaluate. Build each input with policyInput.
• On DENY the closed-over handler never runs. The model receives { arcjetDenied: true, reason, message, retryable } as toolResponse.output.
• Default onGuardError: "deny" blocks the tool if Arcjet is unreachable.
• Prefer omitting outputSchema on guarded tools, or verify the schema accepts ArcjetDenialResult / your onDeny shape. A denial is not schema-checked because the wrapper sits outside action().
• multipart: true tools work too. Their denial is returned as the output field of the multipart response, which is the field generate() copies to toolResponse.output.

Step 3: Gate unwrapped / MCP / filesystem tools

import { guardMiddleware } from "@arcjet/guard/genkit/v1";
import { tokenBucket } from "@arcjet/guard";

import { arcjet } from "./arcjet.js";

const mcpLimit = tokenBucket({
  refillRate: 20,
  intervalSeconds: 60,
  maxTokens: 20,
});

await ai.generate({
  prompt: userText,
  tools: [lookupOrder, ...mcpTools],
  use: [
    guardMiddleware(arcjet, {
      action: ({ toolName }) => `${toolName}.invoked`,
      rules: ({ toolName }) => [mcpLimit({ key: toolName, requested: 1 })],
      sessionId: conversationId,
    }),
  ],
  context: { sessionId: conversationId },
});

Already-branded (guardTool) actions skip the middleware guard when they can be looked up on the registry. Tools that cannot be looked up are still gated.

generate({ context }) is delivered to authored handlers via ALS. The tool-hook ctx from toRunOptions is only { metadata, resumed } — no ALS context. Put the same id on policy.sessionId when you need tool-time correlation through the hook.

Step 4: Screen inbound before generate

import { detectPromptInjection } from "@arcjet/guard";
import { genkitContext } from "@arcjet/guard/genkit/v1";

import { arcjet } from "./arcjet.js";

const appContext = { sessionId: conversationId };
const inbound = detectPromptInjection();
const decision = await arcjet.guard({
  label: "message.received",
  rules: [inbound(userText)],
  ...genkitContext({ context: appContext }),
});

if (decision.conclusion === "DENY") {
  throw new Error("message blocked");
}

await ai.generate({
  prompt: userText,
  tools: [lookupOrder],
  use: [guardMiddleware(arcjet, { sessionId: conversationId })],
  context: appContext,
});

There is no guardInbound.

Step 5: Correlation

Put the id you already have on the app context you pass to generate():

const appContext = { sessionId: conversationId };
await ai.generate({
  prompt: userText,
  context: appContext,
});

Preference order: context.correlationId, then context.sessionId, then context.conversationId, then a caller-owned flowId / runId, then the envelope copies, then init.sessionId. If none is a valid 1–256 printable-ASCII string, the call is uncorrelated rather than joined to a generated id nobody has.

Never read Session.sessionId from a Session constructed without an id — that class mints a UUID. Never read traceId. Never treat interrupt / resumed as correlation.

Verify the integration

1. npm run typecheck passes.
2. Exercise inbound PI (before generate), a tool deny, PII on args, a rate limit, a middleware deny on an unwrapped tool, and fail-closed (an unreachable guard). Confirm finishReason !== "interrupted".
3. Confirm in the Arcjet dashboard that decisions share the session / conversation id as their correlation id.
4. Manual E2E with a real ARCJET_KEY is still-to-verify until you run it.

A full working demo will land in arcjet/examples genkit-agent as a later follow-up. Do not add an example under examples/ in the JS SDK repo.

Note: capture events are fire-and-forget and batched, so events can lag the decisions they accompany by a few seconds. A dropped event is diagnosed, never thrown.