@clawdstrike/vercel-ai

Minimal, runtime-optional wrappers for Vercel AI SDK-style tools.

See Enforcement Tiers & Integration Contract for what is enforceable at the tool boundary (and what requires a sandbox/broker).

Install

npm install @clawdstrike/vercel-ai @clawdstrike/engine-local ai

Usage (tool wrapping)

import { createStrikeCell } from '@clawdstrike/engine-local';
import { createVercelAiInterceptor, secureTools } from '@clawdstrike/vercel-ai';

const engine = createStrikeCell({ policyRef: 'default' });
const interceptor = createVercelAiInterceptor(engine, { blockOnViolation: true });

const tools = secureTools(
  {
    bash: {
      async execute(input: { cmd: string }) {
        return `ran: ${input.cmd}`;
      },
    },
  },
  interceptor,
);

await tools.bash.execute({ cmd: 'echo hello' });

Middleware-style API

import { createStrikeCell } from '@clawdstrike/engine-local';
import { createClawdstrikeMiddleware } from '@clawdstrike/vercel-ai';

const engine = createStrikeCell({ policyRef: 'default' });
const security = createClawdstrikeMiddleware({
  engine,
  config: { blockOnViolation: true, injectPolicyCheckTool: true },
});

const tools = security.wrapTools({
  bash: { async execute(input: { cmd: string }) { return input.cmd; } },
});

Model wrapper (AI SDK)

import { openai } from '@ai-sdk/openai';

const model = security.wrapLanguageModel(openai('gpt-4o-mini'));

Prompt Security (P1)

Enable prompt-security features (instruction hierarchy, jailbreak detection, output sanitization, optional watermarking) for model calls:

import { createStrikeCell } from '@clawdstrike/engine-local';
import { createClawdstrikeMiddleware } from '@clawdstrike/vercel-ai';
import { openai } from '@ai-sdk/openai';

const engine = createStrikeCell({ policyRef: 'default' });
const security = createClawdstrikeMiddleware({
  engine,
  config: {
    blockOnViolation: true,
    // Tool-call streaming evaluation/annotation (best-effort; depends on AI SDK stream part shapes):
    streamingEvaluation: true,
    promptSecurity: {
      enabled: true,
      mode: 'block', // 'warn' | 'audit'
      applicationId: 'my-app',
      jailbreakDetection: { enabled: true },
      instructionHierarchy: { enabled: true },
      outputSanitization: { enabled: true },
      // Optional: embed a signed watermark marker in a system message:
      watermarking: { enabled: true },
    },
  },
});

const model = security.wrapLanguageModel(openai('gpt-4o-mini'));

Notes:

• Prompt-security blocks throw ClawdstrikePromptSecurityError (no raw prompt contents included).
• Prompt-security findings are recorded in security.getAuditLog() as prompt_security_* audit events.

Errors

Blocked tool calls throw ClawdstrikeBlockedError (includes decision and toolName).

React (./react subpath)

The package exports a ./react subpath with React hooks for client-side security integration:

import { useSecureChat } from '@clawdstrike/vercel-ai/react';

const chat = useSecureChat({
  engine,
  // All standard useChat options are supported
  api: '/api/chat',
});

// Security-aware chat state
chat.securityStatus.blocked;      // boolean
chat.securityStatus.checkCount;   // number
chat.securityStatus.violationCount;
chat.blockedTools;                 // string[]
chat.lastDecision;                 // Decision | null
chat.clearBlockedTools();          // reset blocked tool list
await chat.preflightCheck('bash', { cmd: 'rm -rf /' }); // manual preflight

Peer dependencies: @ai-sdk/react, react (both optional).