Vercel AI SDK

Integrate Intutic governance with the Vercel AI SDK (ai, v6+).

The Vercel AI SDK is governed on two independent surfaces, and — unlike almost every other harness in this catalog — only one of them can be wired up without touching your code:

1. Local tool execution — SDK-side, via toolApproval. Wire this into generateText/streamText/ToolLoopAgent and it governs every tool call.
2. LLM egress — in-code only. Read the limitation below before assuming this "just works" the way it does for every other harness.

Positioning vs. @ai-sdk/policy-opa

Vercel ships its own first-party toolApproval adapter, @ai-sdk/policy-opa (confirmed at v1.0.68, Apache-2.0). It evaluates Rego policies via Open Policy Agent — either in-process WASM or a running OPA HTTP server — and maps the decision onto the same ToolApprovalStatus shape toolApproval already understands, with a shadow() wrapper for rolling out a new policy in observe-only mode before enforcing it.

The two are adjacent, not competing: both attach a policy decision to an ai-package tool call. The difference is scope. @ai-sdk/policy-opa is a policy backend for the single ai package — it authors rules in Rego and evaluates them via OPA, with no reach outside a toolApproval callback. Intutic is a cross-framework, cross-harness governance layer: the same proxy, DLP scanner, budget ledger, and audit trail that gates this SDK's LLM egress and tool calls also gate the other 38 harnesses in this catalog (Claude Code, Cursor, LangChain, Strands, ...) from one control plane, with policies as plain files in git rather than compiled Rego bundles. A team already invested in OPA/Rego elsewhere in its stack may reasonably prefer @ai-sdk/policy-opa for ai-package-only tool gating; a team that needs one governance layer across an entire engineering org's tool sprawl — not just the ai package — is what Intutic is built for. Nothing prevents running both: @intutic/gate/vercel's intuticToolApproval() composes with any other toolApproval callback the same way it composes with a human-in-the-loop 'user-approval' status (see step 3 below).

Known, plain limitation: no environment-variable LLM routing

Every other harness in this codebase that governs LLM egress does so by honouring OPENAI_BASE_URL/ANTHROPIC_BASE_URL (or an equivalent), which .env.intutic/intutic exec sets for you with zero code changes. The Vercel AI SDK has no such mechanism. Its providers (@ai-sdk/openai, @ai-sdk/anthropic, etc.) are constructed directly in your code, and the base URL they call is a constructor option, not an environment variable the SDK reads on its own.

This means: .env.intutic's ANTHROPIC_BASE_URL/OPENAI_BASE_URL/INTUTIC_PROXY_URL vars are written (a workspace may run other harnesses that DO honour them, and you're free to read them yourself), but sourcing that file does nothing to route Vercel AI SDK's own LLM calls through the Intutic proxy. Routing requires in-code provider construction — every place your application constructs a model provider must point it at the proxy explicitly. withIntuticProxy() (below) is a small wrapper for that; it is not a way to avoid it.

Setup

1. Initialize Intutic

intutic init

✓ Detected harnesses:
  • vercel-ai-sdk → .env.intutic

Detection requires ai at major version 6 or above (the toolApproval API this integration relies on is a v6+ surface) and at least one @ai-sdk/* provider package — ai alone declares the tool-loop surface but ships no model provider.

2. Route LLM traffic through the proxy (in code — this is the only way)

npm install @intutic/gate

import { createOpenAI } from '@ai-sdk/openai'
import { withIntuticProxy } from '@intutic/gate/vercel'

// Every provider construction call site in your app needs this —
// there is no ambient env var that achieves the same effect.
const openai = withIntuticProxy(createOpenAI)({ apiKey: process.env.OPENAI_API_KEY })
const model = openai('gpt-5-mini')

withIntuticProxy() resolves the proxy URL the same way every other harness's env-writer does (INTUTIC_PROXY_URL, falling back to http://localhost:4000) and passes it as baseURL to whatever provider factory you give it — createOpenAI, createGateway, or any compatible @ai-sdk/* create* export.

3. Gate local tool execution (SDK)

The ai package exposes a documented pre-execution veto point — a toolApproval callback accepted by generateText, streamText, and ToolLoopAgent — which intuticToolApproval() implements:

import { generateText } from 'ai'
import { Gate, install } from '@intutic/gate'
import { intuticToolApproval } from '@intutic/gate/vercel'

install(new Gate({ workspaceId: process.env.INTUTIC_WORKSPACE_ID }))

await generateText({
  model,
  tools,
  toolApproval: intuticToolApproval(),
})

On allow, this resolves to 'not-applicable' — this gate has no opinion, and any other approval mechanism you've configured (e.g. a human-in-the-loop 'user-approval' status for a specific tool) still applies. On deny, it resolves to { type: 'denied', reason }, where reason is the [Intutic Governance] BLOCKED: ... message the SDK hands back to the model in place of a tool result — the tool's real implementation never runs.

4. Trace attribution

import { intuticHeaders } from '@intutic/gate'

const openai = withIntuticProxy(createOpenAI)({
  apiKey: process.env.OPENAI_API_KEY,
  headers: intuticHeaders({ sessionId: runId, harness: 'vercel-ai-sdk' }),
})

What gets written

Same .env.intutic shape as every other SDK-gated framework — proxy URLs plus a pointer at @intutic/gate/vercel. Unlike the Python-SDK-gated family (LangChain, LangGraph, ...), sourcing this file does not route this framework's own LLM egress — see the limitation above.

What the adapter does NOT do

Same structural gaps as every SDK-gated framework — see LangGraph's "What the adapter does NOT do". In short: there is no sync-daemon hook file, argument-level gating requires wiring the SDK into your own agent code, and x-intutic-harness attribution is client-supplied, not authorization. On top of that structural family, this integration is not "zero-code" for LLM-egress routing the way the rest of the catalog is — see the limitation above, stated plainly rather than oversold.

Config details

Property	Value
Harness type	vercel-ai-sdk
Config file	.env.intutic
Detection	ai at major version ≥ 6, plus any @ai-sdk/* package, in package.json (dependencies, devDependencies, or peerDependencies)
Format	Shell environment variables (LLM-egress vars are inert for this framework unless read explicitly — see above)
Write strategy	Atomic (write to .intutic-tmp, then rename)
Tool gate	SDK-side (@intutic/gate/vercel's intuticToolApproval(), the ai package's toolApproval option) — no sync-daemon hook file
LLM-egress routing	In-code only (@intutic/gate/vercel's withIntuticProxy()) — no environment-variable override exists for this framework