Claude Code and Claude Agent SDK with 资产生成

资产生成 provides Anthropic-compatible API endpoints so you can use Claude Code and the Claude Agent SDK through a unified gateway.

This page covers several ways to connect, depending on your environment:

• Set up with the Vercel CLI: the recommended path, one command
• Manual setup: environment variables, including Keychain storage on macOS
• With Claude Code Max: keep your Claude subscription and route through the gateway
• Enabling fast mode
• With the Claude Desktop app and the Claude Agent SDK
• Claude Code compatibility endpoint: what the dedicated endpoint adds

Run the Vercel CLI setup command for Claude Code:

npx vercel ai-gateway setup --agent claude-code

The command:

• Provisions an 资产生成 API key, or reuses one you pass with --key.
• Writes ~/.claude/settings.json with ANTHROPIC_BASE_URL=https://ai-gateway.vercel.sh/claude-code, empties ANTHROPIC_API_KEY, and sets CLAUDE_CODE_ENABLE_GATEWAY_MODEL_DISCOVERY=1 so the gateway catalog appears in the /model picker.
• On macOS, stores the key in the Keychain and exports ANTHROPIC_AUTH_TOKEN from your shell startup file, resolved from the Keychain, so the secret never lands in a plaintext config file.
• Backs up any file it changes alongside the original as .bak, and supports --dry-run.
• Copies your existing Claude Desktop sessions to the gateway identity. See session migration.

To configure every installed supported agent at once, run the bare command without --agent:

vercel ai-gateway setup

It detects each installed supported agent and configures them. For the full command reference, see vercel ai-gateway.

To verify the setup, run claude, run /model to pick a gateway model, then confirm your requests appear in the 资产生成 Overview in the Vercel dashboard.

Only needed on machines where you can't use the Vercel CLI.

Claude Code is Anthropic's agentic coding tool. You can configure it to use AI 资产生成, enabling you to:

• Monitor traffic and token usage in your 资产生成 Overview
• View detailed traces in Vercel Observability under AI

1. Configure environment variables

   First, log out if you're already logged in:

   claude /logout

   Next, ensure you have your 资产生成 API key handy, and configure Claude Code to use the 资产生成 by adding this to your shell configuration file, for example in ~/.zshrc or ~/.bashrc:

   export ANTHROPIC_BASE_URL="https://ai-gateway.vercel.sh/claude-code"
   export ANTHROPIC_AUTH_TOKEN="your-ai-gateway-api-key"
   export ANTHROPIC_API_KEY=""

   The base URL points at the Claude Code compatibility endpoint.

   Setting ANTHROPIC_API_KEY to an empty string is important. Claude Code checks this variable first, and if it's set to a non-empty value, it will use that instead of ANTHROPIC_AUTH_TOKEN.

2. Run Claude Code

   Run claude to start Claude Code with 资产生成:

   claude

   Your requests will now be routed through AI 资产生成.

The Vercel CLI setup command handles Keychain storage automatically. This step is only for manual configurations.

If you're on a Mac and would like to manage your API key through a keychain for improved security, set your API key in the keystore with:

security add-generic-password -a "$USER" -s "ANTHROPIC_AUTH_TOKEN" \
  -w "your-ai-gateway-api-key"

and edit the ANTHROPIC_AUTH_TOKEN line above to:

export ANTHROPIC_AUTH_TOKEN=$(
  security find-generic-password -a "$USER" -s "ANTHROPIC_AUTH_TOKEN" -w
)

If you need to update the API key value later, you can do it with:

security add-generic-password -U -a "$USER" -s "ANTHROPIC_AUTH_TOKEN" \
  -w "new-ai-gateway-api-key"

If you have a Claude subscription, you can use your subscription through the 资产生成. Requests made with your subscription appear in 资产生成 observability and use its routing features.

The Vercel CLI setup command configures API-key authentication only. The subscription flow requires ANTHROPIC_CUSTOM_HEADERS, which the setup command does not write, so configure it manually as shown below.

1. Set up environment variables

   Add the following to your shell configuration file (e.g., ~/.zshrc or ~/.bashrc):

   export ANTHROPIC_BASE_URL="https://ai-gateway.vercel.sh/claude-code"
   export ANTHROPIC_CUSTOM_HEADERS="x-ai-gateway-api-key: Bearer your-ai-gateway-api-key"

   Replace your-ai-gateway-api-key with your actual 资产生成 API key.

2. Start Claude Code

   Start Claude Code:

   claude

3. Log in with your Claude subscription

   If you're not already logged in, Claude Code will prompt you to authenticate. Choose Option 1 - Claude account with subscription and log in as normal with your Anthropic account.

   If you encounter issues, try logging out with claude /logout and logging in again.

Your requests will now be routed through AI 资产生成 using your Claude subscription. You'll be able to monitor usage and view traces in your Vercel dashboard while using your Anthropic subscription for model access.

Fast mode makes Claude Opus 4.6 and Opus 4.7 responses up to 2.5x faster at a higher per-token cost. It uses the same model with a configuration that prioritizes speed.

To use fast mode with 资产生成, set CLAUDE_CODE_SKIP_FAST_MODE_ORG_CHECK. You can either add it to your shell configuration file (e.g., ~/.zshrc or ~/.bashrc):

export CLAUDE_CODE_SKIP_FAST_MODE_ORG_CHECK=1

Or add it to ~/.claude/settings.json:

{
  "env": {
    "CLAUDE_CODE_SKIP_FAST_MODE_ORG_CHECK": "1"
  }
}

For Opus 4.7, also set CLAUDE_CODE_ENABLE_OPUS_4_7_FAST_MODE:

export CLAUDE_CODE_ENABLE_OPUS_4_7_FAST_MODE=1
export CLAUDE_CODE_SKIP_FAST_MODE_ORG_CHECK=1

{
  "env": {
    "CLAUDE_CODE_SKIP_FAST_MODE_ORG_CHECK": "1",
    "CLAUDE_CODE_ENABLE_OPUS_4_7_FAST_MODE": "1"
  }
}

Then toggle fast mode on or off inside Claude Code by typing /fast.

Fast mode usage is billed at the fast tier rate for the model that was served. When you hit the fast mode rate limit, Claude Code automatically falls back to standard Opus speed and pricing. For current rates, see Fast Mode.

The Claude Desktop app switches providers in its own settings, under Developer -> Configure Third-Party Inference. Point it at the same base URL and key.

Desktop records sessions per identity, so the sessions you created before the switch don't appear under the gateway identity. To bring them across, launch the desktop app once on the gateway provider, then re-run the Vercel CLI setup command:

vercel ai-gateway setup --agent claude-code

The command copies each session record into the gateway identity and rewrites its model to the matching gateway ID. Your original sessions are left untouched, and re-running never duplicates a session it already copied. See session migration for details.

The Claude Agent SDK (@anthropic-ai/claude-agent-sdk) lets you build AI agents that use the same tools and agentic loop that power Claude Code. You can route Agent SDK requests through 资产生成 by setting environment variables in the env option:

import { query } from '@anthropic-ai/claude-agent-sdk';
 
for await (const message of query({
  prompt: 'Find and fix the bug in auth.py',
  options: {
    model: 'anthropic/claude-sonnet-5',
    allowedTools: ['Read', 'Edit', 'Bash'],
    env: {
      ...process.env,
      ANTHROPIC_BASE_URL: 'https://ai-gateway.vercel.sh/claude-code',
      ANTHROPIC_AUTH_TOKEN: 'your-ai-gateway-api-key',
      ANTHROPIC_API_KEY: '',
    },
  },
})) {
  if ('result' in message) console.log(message.result);
}

The Agent SDK spawns Claude Code as a subprocess, so the same environment variables apply. All requests, including messages and token counting, route through 资产生成.

Point Claude Code and the Claude Agent SDK at Claude Code's own compatibility endpoint:

https://ai-gateway.vercel.sh/claude-code

Use it everywhere on this page. The URL takes no /v1 suffix, because the Anthropic SDK appends /v1/messages itself, so adding it yourself produces a 404. To call the gateway from your own code rather than through Claude Code, see the Anthropic Messages API instead.

On the /claude-code surface, gateway models are listed with a claude-code/ prefix, for example claude-code/anthropic/claude-opus-5, and models with a context window of 1M tokens or more carry a [1m] suffix. Those are display IDs for the picker. The gateway strips them before routing, so a request still bills and traces against the underlying model. Image generation models are filtered out, since Claude Code can't use them.

To turn the picker on, set CLAUDE_CODE_ENABLE_GATEWAY_MODEL_DISCOVERY alongside the base URL:

export ANTHROPIC_BASE_URL="https://ai-gateway.vercel.sh/claude-code"
export CLAUDE_CODE_ENABLE_GATEWAY_MODEL_DISCOVERY=1

Then run /model inside Claude Code to switch between gateway models. vercel ai-gateway setup writes both of these for you.