Prism
Getting Started

Prism Docs

Quick answer
Prism is a single-binary local proxy that connects any AI coding agent to any LLM provider. It translates Anthropic Messages, OpenAI Chat Completions and OpenAI Responses in real time, brokers an MCP gateway for Model Context Protocol servers, and includes free unlimited web search. It runs on Windows, macOS and Linux with zero runtime dependencies.

Last updated Reviewed against Prism v0.3.26

Start here

Installation

Windows MSI or portable, macOS DMG, Linux AppImage. No runtime dependencies.

Quick Start

From download to a working proxy, with a provider, a model and one connected agent.

Agent Integrations

One-click setup for Claude Code, Codex, Cursor, Zed and ten more agents.

Web Search

Managed SearXNG plus Exa, Tavily, Brave and Serper, with a fallback chain.

How Prism works

Prism runs as two processes. The tray application hosts the admin web UI and the managed SearXNG instance; it spawns a headless child, prism --serve, which serves the proxy. Requests arrive on 127.0.0.1:11434, are resolved to a provider, translated, and forwarded with the response streamed back event by event.

  Your agents                                  Upstream providers
  ───────────                                  ──────────────────

  Claude Code ─────┐                            ┌───────────────────┐
  (Anthropic)      │                            │   Ollama Cloud    │
                   │                            ├───────────────────┤
  Codex Desktop ───┤    ┌────────────────┐      │   OpenCode Go     │
  (Responses)      │    │  Prism proxy    │      ├───────────────────┤
                   ├───▶│  :11434         │─────▶│  Custom OpenAI-   │
  Cursor ──────────┤    │                 │      │  compatible APIs  │
  (Chat Comp.)     │    │  MCP gateway    │      ├───────────────────┤
                   │    │  /mcp           │      │  Codex (OAuth)    │
  Any MCP client ──┤    └────────────────┘      └───────────────────┘
  (JSON-RPC)       │             │
                   │             ├── Search interception ──▶ SearXNG :8888
  Zed · Pi · OMP ──┘             └── Admin UI + Connect panel  :8765

Inbound endpoints

EndpointProtocolAuth
/v1/messagesAnthropic Messagesx-api-key or Bearer
/v1/chat/completionsOpenAI Chat CompletionsBearer
/v1/responsesOpenAI ResponsesBearer
/mcp and /mcp/<agent>MCP JSON-RPCBearer or x-api-key
/v1/modelsModel discoverynone, by design

There is no inbound Ollama /api/chat route. See API Formats for the full endpoint reference and the authentication matrix.

Providers

ProviderCategoryUpstream protocol
Ollama CloudBuilt in, API keyOpenAI Chat Completions
OpenCode GoBuilt in, API keyOpenAI Chat Completions
Custom providersUnlimited, API keyOpenAI Chat Completions (or Ollama native for a local server)
Codex / ChatGPTOAuth, no API keyOpenAI Responses

Key features

MCP Gateway

One Prism-authenticated endpoint for every MCP server, with brokered OAuth.

Web Search

Managed SearXNG for free search, plus Exa, Tavily, Brave and Serper.

One-Click Agent Setup

Auto-detects installed agents and writes their config for them.

Model Remapping

Aliases, per-tier mappings and provider-per-model routing.

Codex via OAuth

Sign in with your ChatGPT account. No API key required.

Stats Dashboard

Live TPS, token counts, per-client breakdown, persisted locally.

Does Prism send anything to a server?

One anonymous heartbeat per day, and nothing else. It carries six fields — a random id, version, OS, architecture, whether Prism was used that day, and a coarse request bucket — never prompts, responses, file paths or keys. You can turn it off in the Admin UI or with PRISM_ANALYTICS_DISABLED=1. See Telemetry & Privacy.

Frequently asked

What is Prism?

Prism is a single-binary local proxy that sits between your AI coding agents and your LLM providers. Agents send requests to http://127.0.0.1:11434 and Prism translates them into whichever protocol the configured upstream expects, then translates the response back.

Which agents and providers does Prism support?

Prism has one-click setup for 14 agents including Claude Code, Codex, Cursor, Factory Droid, OpenCode, ZCode, Grok Build, Zed, Pi, Oh My Pi, Kimi Code, Prime Agent, Empryo, Hermes and DeepSeek Harness. It reaches Ollama Cloud, OpenCode Go, any custom OpenAI-compatible provider, and OpenAI/Codex via OAuth.

Does Prism need Python or Docker?

No. Prism is a single binary with no runtime dependencies. Python is only used by the optional managed SearXNG instance, which can download its own isolated interpreter if your machine has none.

Related

Installation

Install on Windows, macOS or Linux in one step.

Provider Setup

Connect Ollama Cloud, custom APIs and Codex OAuth.

Agent Integrations

Wire up any of 14 supported coding agents.

Troubleshooting

Ports, agent setup, MCP errors and logs.

Guides on the blog

Task-focused walkthroughs: Claude Code on any provider, MCP servers, search providers, Linux and Windows installs.