CLIENT & SDK CONFIG UTILITY

Claude Code, Codex & Cursor API Config Generator

Pick your AI coding agent or backend SDK, choose an active catalog model ID, and copy a verified configuration file plus a minimal connectivity smoke test.

ExecutionZero login · 100% browser-localProtocol templates checked November 11, 2026

Target client or SDKCatalog model IDEnvironment variable nameStreaming mode

1. Configuration file or SDK snippet

request.example

// ~/.claude/settings.json
{
  "env": {
    "ANTHROPIC_BASE_URL": "https://app.llmfly.ai",
    "ANTHROPIC_AUTH_TOKEN": "YOUR_CLAUDE_GROUP_API_KEY",
    "ANTHROPIC_MODEL": "claude-sonnet-5-5",
    "ANTHROPIC_SMALL_FAST_MODEL": "claude-haiku-5-5"
  }
}

2. Minimal connectivity smoke test

request.example

claude -p "Reply with OK." --model claude-sonnet-5-5

Engineering reference and rollout checks

Protocol differences between Claude Code, Codex CLI, and Cursor

Developer tools do not share a single wire format even when they connect to the same gateway domain. Claude Code speaks Anthropic's native Messages protocol and expects ANTHROPIC_BASE_URL set to https://app.llmfly.ai without a trailing /v1 suffix, paired with a key from the Claude Code group.

By contrast, Codex CLI and newer OpenAI reasoning models (such as gpt-6-astra, gpt-6.1-sol, gpt-6-sol, and gpt-6.1-terra) rely on the OpenAI /v1/responses wire API so reasoning summaries, patches, and shell tool calls survive multi-turn execution. Cursor and standard chat libraries connect to https://app.llmfly.ai/v1 using /v1/chat/completions.

Client / SDK	Base URL	Wire protocol	Config file location
Claude Code	https://app.llmfly.ai	Anthropic Messages API	~/.claude/settings.json
Codex CLI	https://app.llmfly.ai/v1	OpenAI Responses API (responses)	~/.codex/config.toml
Cursor IDE	https://app.llmfly.ai/v1	OpenAI Chat Completions	Cursor Settings -> Models
OpenAI Python / TS SDK	https://app.llmfly.ai/v1	Responses or Chat Completions	Server environment variables

Preventing 401, 403, and 404 errors during first-time setup

Three configuration mistakes cause over 90% of initial connection failures: appending /v1 twice (resulting in /v1/v1/chat/completions 404s), typing a marketing display name like "GPT-6 Astra" instead of the exact lowercase catalog ID gpt-6-astra, or using an API key assigned to a different model group than the requested route.

Keep official vendor credentials and LLMFly AI gateway keys in separate shell profiles or project directories so CLI sessions never bill the wrong account.
Run the non-streaming smoke test shown in step 2 before enabling MCP servers, repository indexing, or background sub-agents.
Issue distinct API keys per developer laptop, CI runner, and staging service so leaked or expired keys can be revoked individually.

Frequently asked questions

Why does Claude Code omit /v1 in ANTHROPIC_BASE_URL?

Claude Code appends /v1/messages internally according to Anthropic's SDK convention. Including /v1 in ANTHROPIC_BASE_URL produces /v1/v1/messages and fails with HTTP 404.

When should Python or TypeScript apps use client.responses.create instead of chat.completions.create?

Use client.responses.create for GPT-6 Astra, GPT-6.1 Sol, GPT-6 Sol, and GPT-6.1 Terra—especially when combining reasoning effort with function calling or built-in Responses tools.

Can I use any API key for both Claude Code and GPT models?

Check your key's group assignment in the LLMFly AI console. Claude Code routes and OpenAI Codex/GPT routes belong to specific billing groups in Model Plaza.

Create a dedicated key for your coding environment

Generate a scoped API key in the console and verify your first CLI or SDK request in under a minute.

Get API Key