The model ID
holo3-1-35b-a3b works in both modes. The examples below use it throughout, so you can switch modes by changing only the base URL.
Hosted mode
Sign in once:~/.holo/.env. Check that the key is available:
HAI_API_KEY through the process environment. When the key comes from the environment, holo whoami may not have a cached identity to print.
To pick a model explicitly, pass its ID:
holo3-122b-a10b is hosted only; there are no local weights. See the Models API for the model cards.
MCP and ACP hosts launch the CLI non-interactively, so they cannot complete browser login. Run holo login in a terminal first, then restart the host so it sees the saved key.
Local mode
Local mode runs Holo3.1 on your own machine or network. Bring up an OpenAI-compatible server, then point the CLI at it.Start a server
Follow Local inference in the Models API docs: it covers which size and precision to pick, memory needs, measured speed per hardware, and the tunedllama-server and vllm serve commands for Apple Silicon and DGX Spark.
Connect the CLI
With the server running, open a new terminal and pass its address with--base-url. Local mode does not require holo login.
- llama.cpp
- vLLM
holo run --fast takes one screenshot per step, skips thinking, and uploads smaller images. Expect lower accuracy on fiddly UIs; keep it for simple, well-known flows.
Local mode from hosts
MCP and ACP hosts start the CLI over stdio and read model settings from the environment that launched the host. Set the server URL before the host starts:HAI_AGENT_RUNTIME_BASE_URL is set, MCP and ACP startup does not require HAI_API_KEY.
Shell exports usually do not reach GUI apps launched from the Dock or Finder. If local mode works in your terminal but fails in a host, put both variables in the host’s own MCP or ACP environment config.
Next steps
Local inference
Sizes, precisions, speed, and server commands.
Agent hosts
Wire local mode into an MCP or ACP host.