Claude Code

The Claude Code harness adapter connects HarnessAgent to Claude Code through @anthropic-ai/claude-agent-sdk. The adapter runs a bridge inside the sandbox and streams Claude Code events back to the host over a sandbox-exposed WebSocket.

Harness packages are experimental. Expect breaking changes between releases as this early API gets further refined.

Setup

$ pnpm add @ai-sdk/harness @ai-sdk/harness-claude-code @ai-sdk/sandbox-vercel

The adapter bootstraps the Claude Code bridge dependencies inside the sandbox when the first session starts.

Import

import { claudeCode, createClaudeCode } from '@ai-sdk/harness-claude-code';

claudeCode is equivalent to createClaudeCode() with its default configuration.

Basic Usage

import { HarnessAgent } from '@ai-sdk/harness/agent';
import { claudeCode } from '@ai-sdk/harness-claude-code';
import { createVercelNetworkSandboxSession } from '@ai-sdk/sandbox-vercel';

const agent = new HarnessAgent({
  harness: claudeCode,
  model: 'claude-sonnet-4-6',
});

const sandboxSession = await createVercelNetworkSandboxSession({
  runtime: 'node24',
  ports: [4000],
  template: await agent.getSandboxTemplate(),
});
const session = await agent.createSession({ sandboxSession });

let exitCode = 0;
try {
  const result = await agent.stream({
    session,
    prompt: 'Check the test failures and fix the production code.',
  });

  for await (const part of result.stream) {
    if (part.type === 'text-delta') {
      process.stdout.write(part.text);
    }
  }
} catch (err) {
  exitCode = 1;
  console.error(err);
} finally {
  await session.destroy();
  await sandboxSession.destroy();
  process.exit(exitCode);
}

To use this agent, ensure environment variables include VERCEL_OIDC_TOKEN for Vercel Sandbox, and one of the variables listed under authentication for Claude Code.

Adapter Settings

Use createClaudeCode() to configure the runtime:

const harness = createClaudeCode({
  maxTurns: 10,
  env: {
    DEPLOYMENT_ENV: 'staging',
  },
  thinking: {
    type: 'adaptive',
    display: 'summarized',
  },
});

Settings:

  • auth: authentication mode (auto, direct, or ai-gateway) or an isolated authentication environment.
  • credentialForwarding: optional synchronous or asynchronous callback that customizes each credential immediately before the harness adapter forwards it into a sandbox process. It receives the credential value that would otherwise be forwarded (either the real credential or a masked value) and the environment variable name used to expose it. This callback only controls the value forwarded into the sandbox process. It does not restrict which credentials the harness adapter can discover, read, or otherwise access in the host process.
  • mcpServers: MCP server definitions keyed by server name.
  • strictMcpConfig: only use the servers from mcpServers and ignore other MCP configuration, such as a project .mcp.json or user settings.
  • toolHints: Claude Code hints for host tools, keyed by tool name. readOnly: true marks a tool without side effects, so Claude Code can run its calls in parallel and allows it in plan mode; calls to other host tools run one at a time. alwaysLoad: true sends the tool's definition from the first request when Claude Code's tool search is on, instead of deferring it until the model looks it up.
  • maxTurns: maximum internal turns before yielding.
  • skillOverrides: per-skill visibility overrides. Values can be on, name-only, user-invocable-only, or off.
  • maxBudgetUsd: maximum estimated spend in USD for each Claude query. When it is exceeded, the query stops and the turn ends with an error.
  • questionTimeoutMs: maximum time in milliseconds Claude waits for an answer to an AskUserQuestion question before the question fails. Defaults to the Claude Code default.
  • excludeDynamicSections: moves per-session sections (working directory, memory, git status) out of the Claude Code system prompt into the first user message, so the system prompt can be cached across sessions.
  • agentProgressSummaries: enables periodic AI-generated summaries for running sub-agents. Summaries are available on raw task_progress stream parts.
  • forwardSubagentText: forwards sub-agent text and thinking messages as raw stream parts, in addition to the tool activity forwarded by default.
  • env: environment variables for the Claude Code process. Values are merged over the sandbox bridge process environment and take precedence.
  • thinking: extended-thinking configuration. type can be enabled, disabled, or adaptive. For enabled or adaptive thinking, display can be summarized or omitted. Defaults to { type: 'adaptive', display: 'summarized' }.
  • port: bridge port override.
  • startupTimeoutMs: maximum time to wait for the bridge to start.
  • turnStartTimeoutMs: maximum time from submitting a turn's start to its first bridge event, defaulting to 120 seconds. If no event arrives, the turn fails and the connection closes. This is not a timeout on the complete turn.
  • hostDisconnectTimeoutMs: aborts a running turn when the host disconnects and no host reconnects within this many milliseconds. A suspended turn is not aborted. When unset, a turn keeps running in the sandbox after its host disconnects, including when the host process is gone for good, until the turn finishes or the sandbox stops.
  • reconnect: reconnect timing after an established bridge WebSocket connection drops. maxElapsedMs controls the total retry window, including connection establishment and backoff delays, and defaults to 30 seconds. initialDelayMs defaults to 50 milliseconds, and maxDelayMs defaults to 2 seconds. These retries use exponential backoff and are separate from startupTimeoutMs. They cannot recover when the sandbox, bridge process, or bridge endpoint is permanently unavailable. Pending turn starts are retried with the same identifier when the bridge supports idempotent starts, so a dropped frame does not leave the turn waiting forever or launch the agent twice. An older bridge that is still waiting for a lost start causes the turn to fail instead of retrying unsafely. A turn submitted while disconnected queues its start for first delivery on reconnection; this first delivery also works with older bridges.
  • onDebug: callback for bridge connection lifecycle events: reconnect-attempt, reconnected, and reconnect-failed. reconnect-failed means the connection is closed for good, so a host that caches idle sessions can detach and reattach before sending the next prompt. The cause of a failed reconnect can contain connection details, so avoid logging it as is.
  • mintBridgeToken: synchronous function that receives the sandbox id and returns the bridge authentication token. By default, the adapter generates a random 32-byte token. Custom implementations must return a suitably secret token.

Skill Overrides

Use skillOverrides to control how named skills are exposed to Claude Code without changing shared user-level configuration. Values can be 'on', 'name-only', 'user-invocable-only', or 'off':

const harness = createClaudeCode({
  skillOverrides: {
    'update-config': 'off',
    'fewer-permission-prompts': 'off',
    'keybindings-help': 'off',
    init: 'off',
  },
});

Skills without an override keep their Claude Code defaults, so other bundled skills such as code-review, simplify, and claude-api remain available when provided by the installed Claude Code runtime.

Structured Output

Claude Code supports schema-backed HarnessAgent structured output. The adapter passes the JSON Schema through the Agent SDK's native outputFormat option and returns its structured_output value as JSON text.

Authentication

The auth setting selects how credentials are resolved from the host environment:

  • auto (default): use AI Gateway credentials when available, then fall back to direct Anthropic credentials.
  • direct: use Anthropic credentials.
  • ai-gateway: use AI Gateway credentials.

When the sandbox supports additive request transformations, the bridge receives placeholders and the adapter injects credentials into matching outbound requests. Sandboxes without that capability retain direct credential forwarding.

Supported environment variables:

  • VERCEL_OIDC_TOKEN
  • AI_GATEWAY_API_KEY
  • AI_GATEWAY_BASE_URL
  • ANTHROPIC_API_KEY
  • ANTHROPIC_AUTH_TOKEN
  • ANTHROPIC_BASE_URL

If no applicable credential environment variable is set, the adapter attempts to resolve a native subscription from the host system unless AI Gateway authentication is selected.

Select a specific authentication mode when you do not want automatic detection:

const directHarness = createClaudeCode({ auth: 'direct' });
const gatewayHarness = createClaudeCode({ auth: 'ai-gateway' });

Pass an authentication environment to use programmatically resolved credentials without reading process.env:

const harness = createClaudeCode({
  auth: { ANTHROPIC_API_KEY: await resolveAnthropicToken() },
});

The supplied record replaces the host environment for authentication discovery. Only recognized authentication variables are forwarded.

Sandbox

Claude Code requires a network sandbox with at least one exposed port, e.g. @ai-sdk/sandbox-vercel:

const sandboxSession = await createVercelNetworkSandboxSession({
  runtime: 'node24',
  ports: [4000],
  template: await agent.getSandboxTemplate(),
});

Built-in Tools

The adapter exposes these common Claude Code built-ins through agent.tools:

  • read
  • write
  • edit
  • bash
  • glob
  • grep
  • webSearch

Additional Claude Code built-ins may also appear in agent.tools when they do not fit a common tool shape.

Claude Code supports built-in tool approval requests when permissionMode is allow-reads or allow-edits. A call waiting for approval is kept in the Claude Code session, so an approval answered after a resume runs that same call, even in a new sandbox. Claude Code waits for one approval at a time, so parallel calls that need approval run one after another.

In these modes, file reads are automatically approved only within the session workdir or an active skill's directory. allow-edits automatically approves file changes only within the workdir. Paths outside those directories, including symlink targets, and paths that cannot be resolved require host approval. allow-all remains unrestricted. These approval rules are not a filesystem sandbox, and approved shell commands can access paths outside the workdir.

Known Limitations

Claude Code does not allow toggling whether to discover project-level skills and AGENTS.md files individually from project-level configuration and hooks. Since the adapter disables reading project-level configuration and hooks for security reasons, it means Claude Code will not discover project-level skills and AGENTS.md instructions natively.