Prism 🔺

Universal LLM protocol bridge — use any provider with any coding agent.

What is Prism?

Prism is a proxy that sits between your AI coding tool and any OpenAI-compatible LLM provider. It translates between the Anthropic API format (used by Claude Code, Cline, Roo Code) and the OpenAI API format (used by Cursor, Aider, Continue, Windsurf), so you can use any model from any provider with any coding tool.

┌─────────────┐     ┌─────────────┐     ┌─────────────────┐
│  Claude Code │────▶│             │────▶│  OpenRouter      │
│  (Anthropic) │     │             │     │  Groq            │
├─────────────┤     │   Prism     │     │  Mistral         │
│  Cursor      │────▶│   Proxy     │────▶│  Together AI     │
│  (OpenAI)    │     │             │     │  NVIDIA NIM      │
├─────────────┤     │             │     │  Ollama          │
│  Aider       │────▶│             │────▶│  Any OpenAI-compat│
│  (OpenAI)    │     └─────────────┘     └─────────────────┘
└─────────────┘

Quick Start

# Install
pip install prism-proxy

# Run with interactive model mapping (recommended)
prism --provider https://openrouter.ai/api/v1 --key sk-... --interactive

# Or run with a single model
prism --provider https://api.groq.com/openai/v1 --key sk-... --model llama-3.3-70b-versatile

Connect Your Tool

Claude Code / Cline / Roo Code

export ANTHROPIC_BASE_URL=http://localhost:8000
export ANTHROPIC_API_KEY=prism  # any value works
claude  # or cline, roo-code

Cursor / Aider / Continue / Windsurf

export OPENAI_BASE_URL=http://localhost:8000
export OPENAI_API_KEY=prism  # any value works

In Tool Settings

Tool	Setting	Value
Claude Code	ANTHROPIC_BASE_URL	http://localhost:8000
Cursor	Settings → Models → OpenAI Base URL	http://localhost:8000/v1
Aider	--openai-api-base http://localhost:8000/v1
Cline	API Provider → Custom OpenAI Compatible	http://localhost:8000
Continue	models[].apiBase	http://localhost:8000

Supported Providers

Any OpenAI-compatible endpoint works:

Provider	Base URL	Notes
OpenRouter	https://openrouter.ai/api/v1	Best variety, 200+ models
Groq	https://api.groq.com/openai/v1	Fast inference
Mistral AI	https://api.mistral.ai/v1	Good code models
Together AI	https://api.together.xyz/v1	Open source models
NVIDIA NIM	https://integrate.api.nvidia.com/v1	Enterprise models
Ollama	http://localhost:11434/v1	Local models
vLLM	http://localhost:8000/v1	Self-hosted
LM Studio	http://localhost:1234/v1	Local GUI

Supported Clients

Client	Format	Status	Notes
Claude Code	Anthropic	✅ Full	Streaming, tools, thinking
Cline / Roo Code	Anthropic	✅ Full	Streaming, tools
Cursor	OpenAI	✅ Full	Streaming, tools
Aider	OpenAI	✅ Full	Streaming, tools
Continue	OpenAI	✅ Full	Streaming, tools
Windsurf	OpenAI	✅ Full	Streaming
GitHub Copilot	OpenAI	✅ Basic	Chat completions
Any OpenAI client	OpenAI	✅ Full	Auto-detected
Any Anthropic client	Anthropic	✅ Full	Auto-detected

Features

• Auto-format detection — Works with both Anthropic and OpenAI APIs automatically
• Streaming — Full SSE streaming with proper event sequences
• Tool use — Complete tool/function calling support
• Thinking — Extended thinking / reasoning block support
• Prompt caching — Cache control markers preserved
• Token counting — /v1/messages/count_tokens endpoint
• Model mapping — Route any Claude model to any provider model
• Interactive setup — Fetches provider models and prompts for mappings
• Multi-key round-robin — Distribute load across API keys
• Error translation — Provider errors mapped to client-expected format
• CORS support — Works with browser-based tools
• Health dashboard — /health endpoint with status info

CLI Options

prism [OPTIONS]

Options:
  --provider URL        Provider base URL (required)
  --key KEY             Single API key
  --keys KEYS           Comma-separated keys for round-robin
  --model MODEL         Single model to use
  --model-map MAP       Comma-separated client=provider pairs
  --interactive, -i     Interactive model mapping
  --fallback-model M    Default model when no mapping matches
  --host HOST           Bind address (default: 127.0.0.1)
  --port PORT           Server port (default: 8000)
  --log-level LEVEL     debug/info/warning
  --config, -c FILE     YAML config file

All flags have PRISM_* env var equivalents (e.g. PRISM_PROVIDER, PRISM_KEY).

Config File

Create prism_config.yaml:

provider_url: https://openrouter.ai/api/v1
api_key: sk-or-...
model_map:
  claude-opus-4-6: anthropic/claude-opus-4
  claude-sonnet-4-6: anthropic/claude-sonnet-4
  claude-haiku-4-5: anthropic/claude-3.5-haiku
fallback_model: anthropic/claude-3.5-haiku
host: 127.0.0.1
port: 8000

Run with: prism -c prism_config.yaml

Model Mapping

Map Claude model names to your provider's models:

# Command line
prism --provider https://openrouter.ai/api/v1 --key sk-... \
  --model-map 'claude-opus-4-6=anthropic/claude-opus-4,claude-sonnet-4-6=anthropic/claude-sonnet-4'

# Interactive (recommended)
prism --provider https://openrouter.ai/api/v1 --key sk-... --interactive

Built-in Model Aliases

Prism includes pre-configured mappings for popular providers. When you use a known provider URL, Claude models are automatically mapped:

Claude Model	OpenRouter	Groq	Mistral
claude-opus-4-6	anthropic/claude-opus-4	llama-3.3-70b-versatile	mistral-large-latest
claude-sonnet-4-6	anthropic/claude-sonnet-4	llama-3.1-70b-versatile	codestral-latest
claude-haiku-4-5	anthropic/claude-3.5-haiku	llama-3.1-8b-instant	ministral-8b-latest

API Endpoints

Endpoint	Method	Description
/v1/messages	POST	Anthropic message API
/v1/chat/completions	POST	OpenAI chat completions API
/v1/models	GET	List available models
/v1/messages/count_tokens	POST	Count tokens in messages
/health	GET	Health check & status
/	GET	Setup instructions

What Gets Translated

Requests:

• Content blocks ↔ messages
• System prompts
• Tool definitions / input_schema
• Tool results
• Model names (via mapping)
• Thinking blocks
• Images (base64 and URLs)
• Cache control markers

Responses:

• choices[0].message ↔ content[]
• finish_reason ↔ stop_reason
• Tool calls
• Token usage (including cache tokens)
• Thinking/reasoning blocks

Streaming:

• Text deltas
• Tool call chunks (partial JSON)
• Thinking deltas with signatures
• Proper SSE event format for each protocol

Never touched: Tool function names/arguments, message text, conversation history, auth.

Health Dashboard

curl http://localhost:8000/health

{
  "status": "ok",
  "version": "0.7.0",
  "uptime_seconds": 123.4,
  "requests": 42,
  "errors": 0,
  "provider": "https://openrouter.ai/api/v1",
  "model": "anthropic/claude-sonnet-4",
  "model_map_count": 3,
  "keys_configured": 1,
  "endpoints": {
    "anthropic": ["/v1/messages", "/messages"],
    "openai": ["/v1/chat/completions", "/chat/completions"],
    "models": ["/v1/models", "/models"],
    "token_count": ["/v1/messages/count_tokens"]
  }
}

Install

# From PyPI
pip install prism-proxy

# From source
git clone <url>
cd prism
pip install -e .

License

MIT