@artkit-ai/ai-sdk

Vercel AI SDK 7 adapter for artkit: tools, approval, tool context, data-artifact parts and the useArtifacts hook. Peer dependency: ai@^7 (and react@^19 for /react).

Server

import {
  ToolLoopAgent,
  convertToModelMessages,
  createUIMessageStream,
  createUIMessageStreamResponse,
} from 'ai';
import {
  artifactContextSchema,
  artifactInstructions,
  artifactToolApproval,
  artifactTools,
  artifactToolsContext,
} from '@artkit-ai/ai-sdk';

const tools = artifactTools(toolkit);
const agent = new ToolLoopAgent({
  model,
  instructions: [baseInstructions, artifactInstructions(toolkit)].join('\n\n'),
  tools,
  toolApproval: artifactToolApproval(toolkit), // submit, apply on published… → 'user-approval'
  callOptionsSchema: artifactContextSchema, // { workspaceId, actor, writer? }
  prepareCall: ({ options, ...settings }) => ({
    ...settings,
    toolsContext: artifactToolsContext(tools, options),
  }),
});

export async function POST(request: Request) {
  const { messages } = await request.json();
  const { workspaceId, actor } = await authenticate(request); // never from the body
  const stream = createUIMessageStream({
    execute: async ({ writer }) => {
      const result = await agent.stream({
        messages: await convertToModelMessages(messages, { tools }),
        options: { workspaceId, actor, writer },
      });
      writer.merge(result.toUIMessageStream());
    },
  });
  return createUIMessageStreamResponse({ stream });
}

• Tools keep the toolkit's names, descriptions and JSON Schemas (jsonSchema()); input is validated by the toolkit so the model receives fixable error payloads instead of SDK tool errors.
• Scope comes from the tool context (toolsContext), the tool call id is the idempotency key.
• Each commit writes { type: 'data-artifact', id: artifactId, data: { id, kind, title, version, status } }, updated in place by id: messages carry references, not content.
• For WorkflowAgent (@ai-sdk/workflow), use artifactTools(toolkit, { toolLevelApproval: true }), which sets the tool-level needsApproval that durable workflows use.

Client

import { useChat } from '@ai-sdk/react';
import { lastAssistantMessageIsCompleteWithApprovalResponses } from 'ai';
import { useArtifacts } from '@artkit-ai/ai-sdk/react';

const { messages, addToolApprovalResponse } = useChat({
  sendAutomaticallyWhen: lastAssistantMessageIsCompleteWithApprovalResponses,
});
const { artifacts, byId, pendingApprovals, approve, reject } = useArtifacts(messages, {
  registry,
  addToolApprovalResponse,
});

Each ArtifactState has the last committed reference, a provisional preview while artifact_create or artifact_apply input is streaming (create content goes through the kind's parsePartial; apply keeps complete operations only), the approval state (approval-requested, approved, rejected) and the last tool error. deriveArtifacts(messages, options) is the same logic without React.

AI SDK 7 names

Confirmed against [email protected]: ToolLoopAgent, tool, jsonSchema, per-tool contextSchema + toolsContext map, agent-level toolApproval (statuses user-approval, not-applicable…), callOptionsSchema/prepareCall, createUIMessageStream, UIMessageStreamWriter, tool part states including approval-requested, approval-responded and output-denied, and useChat().addToolApprovalResponse. WorkflowAgent lives in @ai-sdk/workflow and HarnessAgent in @ai-sdk/harness; they emit the same UI stream, so useArtifacts works with them, but they are not dependencies of this package.