Adapters

Claude Code

The Claude Code adapter runs Claude Code (via the @anthropic-ai/claude-agent-sdk) as a chat backend. Unlike HTTP provider adapters, this is a harness adapter: Claude Code runs its own agent loop and executes its own tools — bash, file reads and edits, glob/grep search, web search — locally on your server. Each chat() call runs one full harness turn; the harness's tool activity streams back as already-resolved tool-call events your UI can render.

Server-only. The harness spawns the Claude Code runtime as a subprocess, so this adapter only works in a Node.js server environment — never in the browser. Treat it like giving Claude a shell on the machine it runs on, and configure permissions accordingly.

Installation

shell
npm install @tanstack/ai-claude-code

A runnable demo lives at examples/sandbox-web — switch the harness (Claude Code, Codex, OpenCode, Grok Build) and sandbox provider per run, with session resume, the harness tool timeline, permission modes, and tool bridging, wired into a TanStack Start app.

Authentication

The harness resolves credentials the same way Claude Code does:

  • ANTHROPIC_API_KEY in the server's environment (or the apiKey config option), or

  • an existing Claude subscription login on the machine (claude login).

Basic Usage

ts
import { chat } from "@tanstack/ai";
import { claudeCodeText } from "@tanstack/ai-claude-code";

const stream = chat({
  adapter: claudeCodeText("claude-opus-4-8", {
    cwd: "/path/to/project",
    permissionMode: "acceptEdits",
  }),
  messages: [{ role: "user", content: "Fix the failing test in utils.test.ts" }],
});

Configuration

OptionDescription
cwdWorking directory for the harness session. Defaults to process.cwd().
permissionModeClaude Code permission mode ('default', 'acceptEdits', 'bypassPermissions', 'plan', 'dontAsk', 'auto'). See the permissions note below.
allowedToolsBuilt-in tools the harness may use without prompting (e.g. ['Read', 'Grep', 'Bash(npm test:*)']).
disallowedToolsBuilt-in tools removed from the harness entirely.
maxTurnsMaximum harness-internal turns per run.
systemPromptMode'append' (default) keeps Claude Code's preset system prompt and appends your systemPrompts; 'replace' sends yours as the entire prompt.
mcpServersExtra MCP servers passed through to the harness untouched.
apiKeyAnthropic API key for the harness subprocess.
envExtra environment variables for the harness subprocess.
pathToClaudeCodeExecutableUse a specific Claude Code executable instead of the SDK's bundled one.
streamPartialsEmit true token-level text deltas (default true).
canUseToolCustom permission handler; replaces the adapter's default handler.
settingSourcesClaude Code settings tiers to load. Default ['project']: the cwd's CLAUDE.md and project settings apply, but user-level config on the host (~/.claude plugins, hooks, skills) is ignored. Pass ['user', 'project', 'local'] for CLI-equivalent behavior, or [] for full isolation.

Permissions on headless servers. Without an explicit permissionMode or canUseTool, the adapter installs a safe default handler: bridged TanStack tools always run, and any built-in tool call that would normally prompt a human is denied with guidance instead of hanging the request. To let the harness edit files or run commands, set permissionMode: 'acceptEdits' / 'bypassPermissions', or enumerate allowedTools.

Stateful Sessions

Claude Code sessions are stateful — the harness keeps the full working context (files read, commands run, conclusions reached) between turns. The adapter surfaces the session id of every run as a custom stream event named claude-code.session-id; thread it back via modelOptions.sessionId to resume the session. When resuming, only the latest user message is sent — the harness already holds the prior context.

Server endpoint:

ts
import {
  chat,
  chatParamsFromRequest,
  toServerSentEventsResponse,
} from "@tanstack/ai";
import { claudeCodeText } from "@tanstack/ai-claude-code";

export async function POST(request: Request) {
  const params = await chatParamsFromRequest(request);

  // Extra fields the client puts in the connection `body` arrive here.
  const sessionId =
    typeof params.forwardedProps.sessionId === "string"
      ? params.forwardedProps.sessionId
      : undefined;

  const stream = chat({
    adapter: claudeCodeText("claude-opus-4-8", {
      cwd: "/path/to/project",
      permissionMode: "acceptEdits",
    }),
    messages: params.messages,
    modelOptions: { sessionId },
  });

  return toServerSentEventsResponse(stream);
}

Client (React) — capture the session id from the custom event and send it back on subsequent requests:

ts
import { useState } from "react";
import { useChat } from "@tanstack/ai-react";
import { fetchServerSentEvents } from "@tanstack/ai-client";

function CodingAssistant() {
  const [sessionId, setSessionId] = useState<string | undefined>(undefined);

  const { messages, sendMessage } = useChat({
    connection: fetchServerSentEvents("/api/chat", () => ({
      body: { sessionId },
    })),
    onCustomEvent: (name, value) => {
      if (
        name === "claude-code.session-id" &&
        typeof value === "object" &&
        value !== null &&
        "sessionId" in value &&
        typeof value.sessionId === "string"
      ) {
        setSessionId(value.sessionId);
      }
    },
  });

  // ... render messages; harness tool activity (Bash, Edit, Read, ...)
  // arrives as regular tool-call parts with their results attached.
}

Sessions are stored on the machine that ran them (~/.claude/projects/), so resuming only works on the same server instance. Pass modelOptions: { forkSession: true } alongside sessionId to branch a session instead of continuing it.

Tools

Two kinds of tools flow through this adapter:

  1. Built-in harness tools (Bash, Read, Write, Edit, Glob, Grep, WebSearch, ...) are executed by Claude Code itself. Their activity streams back as tool-call events with results already attached, so useChat UIs render them with no extra wiring — but your code never executes them.

  2. Your TanStack tools are bridged into the harness as an in-process MCP server. Define them as usual with toolDefinition().server(); the model sees them as mcp__tanstack__<name> and the adapter strips the prefix on the way back out, so events match the names you registered.

ts
import { z } from "zod";
import { chat, toolDefinition } from "@tanstack/ai";
import { claudeCodeText } from "@tanstack/ai-claude-code";

const lookupTicket = toolDefinition({
  name: "lookup_ticket",
  description: "Look up an issue ticket by id",
  inputSchema: z.object({ ticketId: z.string() }),
}).server(async ({ ticketId }) => {
  return { ticketId, status: "open", title: "Crash on startup" };
});

const stream = chat({
  adapter: claudeCodeText("claude-opus-4-8"),
  messages: [{ role: "user", content: "What's the status of ticket T-123?" }],
  tools: [lookupTicket],
});

Client-side and approval-gated tools are not supported. The harness executes tools inside a live subprocess, which cannot pause across HTTP requests to wait for a browser round-trip or a human approval. Passing a tool without a server execute() implementation — or one marked needsApproval — fails fast with a descriptive error. Run those tools outside the harness with a regular provider adapter.

Structured Output

structuredOutput() uses the harness's native JSON-schema output format in a one-shot run (single turn, no tools). It works for finalization after a chat, but a plain provider adapter (e.g. @tanstack/ai-anthropic) is the better choice when structured extraction is the primary job — it's faster and doesn't spawn a subprocess.

Limitations

  • Server-only (Node). The harness spawns a subprocess; Windows support is untested.
  • The harness owns the agent loop. TanStack's agent-loop strategies and per-iteration middleware don't apply inside a harness turn; maxTurns is the equivalent control.
  • No sampling controls. temperature-style options don't exist here.
  • Sessions are machine-local. Resume requires hitting the same server instance.
  • Cold starts. Each call spawns a harness turn; expect higher first-token latency than HTTP adapters.