Start here

The gateway is api.conifer.build. Point an OpenAI or Anthropic client at it with one key and every model in the catalog is behind it, at the prices the catalog publishes. Name a model and that model answers. Leave the choice to Conifer and it picks one. Either way the gateway chooses the provider, retries elsewhere if one fails, and answers with a receipt saying which model served and what it cost. By the end of this page you will have a key, a first request, and your tools pointed at the gateway.

How it works#

1. Mint a key

   Open conifer.build/console#/keys and mint a key. It is free to hold and spends nothing until it is used. Save it as CONIFER_API_KEY. Every tool and client in these docs reads that one name.

   terminal

   export CONIFER_API_KEY='sk-conifer-…'

2. Point a client at the gateway

   Set two variables and the official OpenAI package works unchanged against https://api.conifer.build. The Anthropic and Responses wires are on the same host with the same key, so use whichever wire your client already speaks.

   terminal

   export OPENAI_BASE_URL=https://api.conifer.build/v1
   export OPENAI_API_KEY=$CONIFER_API_KEY

   Or try the wire with nothing installed.

   terminalcopy

   curl https://api.conifer.build/v1/chat/completions \1
     -H "Authorization: Bearer $CONIFER_API_KEY" \2
     -H "Content-Type: application/json" \
     -d '{
       "model": "claude-haiku-4-5",3
       "max_tokens": 1024,
       "messages": [{"role": "user", "content": "three names for a build cache"}]
     }'

   1. 1The OpenAI wire. The Anthropic wire is /v1/messages on the same host.
   2. 2Your key, sent as the bearer token. One key covers every model.
   3. 3The id of any chat model in the catalog, from a frontier lab or an open-weights one. It is served as named.

3. Wire your tools

   Claude Code, Codex, Cursor and VS Code each take a base URL and the key. Name a model and that model answers. Send auto and Conifer picks one for you. For every request the gateway chooses the provider, fails over if that provider is down, manages the prompt cache, and answers with a receipt naming the model and the cost. Routing has the rest.

Pick your road#

Let your agent do it#

If you use Claude Code, Codex, Cursor, or another coding agent, paste this prompt and let it do the setup.

Fetch https://conifer.build/llms.txt and follow it: set up Conifer in this project.

/llms.txt is the whole setup in one file, written for a model to read.

Put it inside your editor#

Claude Code, Codex, Cursor, and VS Code each have a page with the config to paste. Everything else covers the other terminal agents. The MCP server lets an agent you already run call any model on the gateway mid-task, with the cost of each call in the result. Coming from OpenRouter, Vercel AI Gateway, or Helicone, migrating takes two lines.

Three things people hunt for#

The key. Save it as CONIFER_API_KEY. The official OpenAI and Anthropic packages take the same key through their own variables, as in step two. The API key covers minting and the SDK covers calling.

Named vs routed. Name a catalog id and that model answers, or the request is refused with a clear error. Send auto and Conifer chooses. The receipt headers name the id you sent and the model that served, so you can always check. Routing has the details, including failover.

When something breaks. 401 is a bad key, 402 is an empty balance, 404 is a model id the catalog does not carry. The full list is on Errors. For anything else, write to us.

The map#

The CLI is optional. It installs with one command and signs in with the same key. On Apple Silicon and Windows it also carries its own engine, so it can run an open-weights model on your hardware. On Linux it is a thin client that talks to the gateway. Local covers that in four short pages.

curl -fsSL https://conifer.build/setup | bash