Clients

Clients

Connect Claude Code, Codex, OpenCode, Grok Build, Hermes, Raycast, ChatGPT for Mac or DeepSeek Harness to OpenLLM, or any OpenAI or Anthropic client.

Each supported client has a command that connects it to OpenLLM. Most run per session and leave your own config alone; Raycast, the ChatGPT Mac app, and Hermes profiles are persistent and come with a command to undo them.

ClientCommandHow it connectsGuide
Claude Codeopenllm claudePer session: run-local --settings and --mcp-configClaude Code
Codexopenllm codexPer session: -c overrides on your configCodex
OpenCodeopenllm opencodePer session: merged config file via OPENCODE_CONFIGOpenCode
Grok Buildopenllm grokPer session: private GROK_HOME over your ~/.grokGrok Build
Hermesopenllm hermesPer session, or a sticky profile with openllm hermes installHermes
Raycastopenllm raycastPersistent: provider block in providers.yamlRaycast
ChatGPT for Macopenllm chatgptPersistent: an OpenLLM ChatGPT launcher appChatGPT for Mac
DeepSeek Harnessdsh plugin --profile web add github:openllmsh/dshPersistent: a bundle in a dsh profile, through the local daemonDeepSeek Harness

The openllm commands need the OpenLLM CLI. See Install.

How the launchers work

  • The gateway is chosen at launch. If your local daemon answers on http://127.0.0.1:8787, the client talks to it. Otherwise it talks to https://www.openllm.sh. Persistent setups (Raycast and the Hermes profile) keep the address from when you ran the command.

  • Arguments pass through. Everything after the client name goes to the client unchanged, so openllm claude --resume behaves like claude --resume.

  • OpenLLM's flags go first. Put them before the client name:

    FlagWhat it does
    -dAdds the client's own skip-approvals flag. Fails on clients that have none.
    --newAlways starts a new session instead of offering to attach to a running one.
    --attach [id]Attaches to a running session on this machine without asking.

  • Models map to aliases. Each client defaults to ultra, and plus and lite are available everywhere. Most clients also list every model you have activated; OpenLLM configures only the three aliases for Hermes and DeepSeek Harness. Change what the aliases mean in Configuration. See Routing and fallback.

  • Tools come along. Every launcher adds the OpenLLM MCP server, except Raycast. See MCP server.

The gateway is chosen at each launch; everything after the client name goes to the client.

Who pays for a request

Every client sends requests to OpenLLM with your sk-llm key. Which account pays depends on the model that serves the request:

ModelWho pays
A subscription model, such as Claude Code or Codex on your own planYour subscription, through the daemon on one of your devices
An API-key modelYour account with that provider
A local model, such as OllamaRuns on your own machine, through the daemon
An included openllm/ modelOpenLLM, up to your plan's monthly allowance

Aliases such as ultra are fallback chains, so one request can try more than one of these. On the Free plan, a monthly token cap also applies. See Providers, Included models, and Plans and limits.

Don't see your tool?

Anything that speaks the OpenAI or Anthropic API can use OpenLLM with a base URL and your key. See Any OpenAI or Anthropic client. For MCP-only tools, see MCP server.

Last updated on

Install OpenLLM

Install the openllm CLI and daemon on macOS or Linux with one curl command, or install the Mac app, then check, update, or uninstall OpenLLM.

Claude Code

Run Claude Code through OpenLLM with openllm claude, map the Opus, Sonnet and Haiku slots to your models, or set ANTHROPIC_BASE_URL yourself.

On this page

How the launchers workWho pays for a requestDon't see your tool?