How to Use Any LLM Provider with Claude Code and Prism

Claude Code has a hard limitation: it only knows how to talk to the Anthropic Messages API. That means you are locked to Anthropic's models for the main thread, for quick completions and for every background sub-agent it spawns. If you want a cheaper daily driver, a faster-inference model, or a provider with different rate limits, Claude Code cannot do it on its own.

Prism solves this by acting as a local proxy between Claude Code and your upstream provider. When Claude Code sends an Anthropic Messages request, Prism translates it to the protocol that provider actually speaks — OpenAI Chat Completions or OpenAI Responses — and streams the response back in Anthropic's shape. Claude Code never knows the difference.

How it works

  Claude Code
       │  Anthropic Messages → http://127.0.0.1:11434
       ▼
  ┌──────────────────────┐
  │  Prism proxy          │  Anthropic ⇄ Chat Completions / Responses
  │  /v1/messages         │  per-model provider routing
  └──────────────────────┘
       │  forwards in the provider's native protocol
       ▼
  Ollama Cloud · OpenCode Go · OpenRouter · Groq · DeepSeek · any OpenAI-compatible API

Claude Code sends all requests to one base URL. Point that at Prism instead of api.anthropic.com and every request flows through the proxy. Prism resolves the requested model to a provider, translates, and forwards. Streaming, thinking blocks, tool calls and system prompts all pass through.

Step by step

What Prism writes

{
  "env": {
    "ANTHROPIC_BASE_URL": "http://127.0.0.1:11434",
    "ANTHROPIC_AUTH_TOKEN": "prism",
    "ANTHROPIC_DEFAULT_OPUS_MODEL": "glm-5.1:cloud",
    "ANTHROPIC_DEFAULT_SONNET_MODEL": "deepseek-v4-flash:cloud",
    "ANTHROPIC_DEFAULT_HAIKU_MODEL": "deepseek-v4-flash:cloud",
    "CLAUDE_CODE_SUBAGENT_MODEL": "deepseek-v4-flash:cloud"
  }
}

Claude Code watches and rewrites ~/.claude/settings.json itself, so those keys can disappear. Prism re-writes them on every startup, which means restarting Prism recovers the integration.

Per-tier model mapping

Claude Code uses different models for different jobs: a strong model for the main thread, a fast one for quick completions, and a lightweight one for the sub-agents it spawns in the background. Prism exposes each as its own environment variable, so all four can point at different upstream models.

ANTHROPIC_DEFAULT_OPUS_MODEL   → main thread (most capable)
ANTHROPIC_DEFAULT_SONNET_MODEL → balanced default
ANTHROPIC_DEFAULT_HAIKU_MODEL  → fast background completions
CLAUDE_CODE_SUBAGENT_MODEL     → parallel sub-agents

This is the single biggest cost lever in a Claude Code setup, because background tasks usually outnumber the messages you type. Point haiku and subagent at a small fast model and the main thread at your strongest one.

Aliasing model names

If a client sends an Anthropic model name that is not in your catalog, Prism serves your default model. To control that mapping explicitly, add an alias under the aliases key in model_remapping.json:

{
  "default_model": "glm-5.1:cloud",
  "known_models": [
    { "id": "glm-5.1:cloud", "provider": "ollama_cloud" }
  ],
  "aliases": {
    "claude-3-5-sonnet-20241022": "glm-5.1:cloud",
    "claude-3-5-haiku-20241022": "deepseek-v4-flash:cloud"
  }
}

When Claude Code asks for claude-3-5-sonnet-20241022, Prism swaps the name before forwarding and the client still sees its original model name in the response. See Model Remapping for the resolution order.

Supported providers

Anything OpenAI-compatible works, plus the two built-ins and Codex via OAuth:

Get started

Install Prism, add a provider, add your models, and click Setup for Claude Code. In a few minutes you can be running it on models from Ollama Cloud, OpenCode Go, OpenRouter, Groq or DeepSeek — with no patches, no forks and no changes to Claude Code itself.

Download Prism