Getting Started
Quick Start
Quick answer
Run
prism.exe (or the AppImage / app bundle), add a provider and one or more models in the Admin UI, then click Setup next to your coding agent in the Agents tab. The agent is now talking to Prism, and Prism is talking to whichever upstream providers you configured.Last updated Reviewed against Prism v0.3.26
Step by step
1
Download Prism
Take the per-user MSI or portable
prism.exe on Windows, the DMG on macOS, or the AppImage on Linux, from the latest release. See Installation for artifact names.2
Run it
Launch the app or the binary. A native tray icon appears and the proxy starts listening on
127.0.0.1:11434. There is nothing to install first and no runtime to configure.# Windows .\prism.exe # macOS open /Applications/Prism.app # Linux ./Prism-Linux-x86_64.AppImage
3
Add a provider
Open
http://127.0.0.1:8765/admin and go to the Provider tab. Choose Ollama Cloud or OpenCode Go, add a custom OpenAI-compatible provider such as OpenRouter or Groq, or add a Codex account through OAuth. See Provider Setup.4
Add your models
In the Models tab, type a model name and click Search. Prism fills in context length, output limits, reasoning support, tool calling, vision and structured outputs from
models.dev, and asks Ollama Cloud directly for its own models. Assign each model a provider and set the default.5
Connect your agent
Go to the Agents tab. Prism lists every supported agent it detects on the machine — Claude Code, Codex, Cursor, Factory Droid, OpenCode, ZCode, Grok Build, Zed, Pi, Oh My Pi, Kimi Code, Prime Agent, Empryo, Hermes and DeepSeek Harness. Click Setup next to one and Prism backs up its config and writes the right entries.
6
Add free web search (optional)
Open the SearXNG tab and click Start. The first start installs an isolated SearXNG runtime (about 80 MB) and then begins listening on
127.0.0.1:8888. Agents that call a web search tool are served locally, with no API key. You can also switch to Exa, Tavily, Brave or Serper in the Search tab.7
Add MCP servers (optional)
Open the MCP tab to browse the registry marketplace or add a server manually, then grant per-agent access. Every agent then reaches MCP through one Prism endpoint,
/mcp, instead of holding its own credentials. See MCP Gateway.Point an agent at Prism manually
One-click setup writes these values for you, but they are all it takes to connect a client by hand.
# Claude Code (Anthropic protocol) ANTHROPIC_BASE_URL = http://127.0.0.1:11434 ANTHROPIC_AUTH_TOKEN = prism # OpenAI SDK / OpenAI-compatible clients OPENAI_BASE_URL = http://127.0.0.1:11434/v1 API key = prism
The proxy key is always prism
Every client authenticates to Prism with the fixed token
prism, sent as Authorization: Bearer prism or x-api-key: prism. Your upstream provider keys live in Prism, never in the agent.Verify it works
Send one request and confirm Prism translates and routes it:
curl http://127.0.0.1:11434/v1/chat/completions \
-H "Content-Type: application/json" \
-H "Authorization: Bearer prism" \
-d '{
"model": "glm-5.1:cloud",
"messages": [{"role": "user", "content": "Hello"}]
}'If you get a completion, the proxy is working. The Stats Dashboard will show the request, the detected client, and the token counts.