Skip to main content
Run OpenHands on your computer and choose a local model server below. For the hosted Nebius path, get your assigned API key through the 1Password access handoff; those model requests go to Nebius. For cloud access through OpenRouter, see the OpenRouter sign-in example.

​
Start OpenHands

Install and start Docker Desktop, then install uv. On Windows, use WSL. See the OpenHands local setup guide for platform requirements. Replace the project path with a disposable copy of your repository. OpenHands will be able to edit files in that directory.
Keep the terminal open, then visit http://localhost:3000 in your browser.

​
Connect local runtimes

OpenHands is not currently listed in Ollama’s launcher integrations, so it requires a connection profile. Start a compatible Flint server using Ollama, Unsloth, or llama.cpp. In Settings → LLM → Advanced, choose None for Provider Connection and create a local profile: Name the profile Flint (local), save, select it as active, and start a new conversation. Here openai/ selects the compatible API protocol; the configured Base URL determines the destination. Verify the model ID with /v1/models. Configure enough server context for OpenHands’ system prompt and tools; its local-model guide recommends 32K or more when hardware permits.

​
Local server networking

The browser’s location does not determine where model requests originate. If the OpenHands process making those requests is in Docker, 127.0.0.1 points inside that container. Use http://host.docker.internal:11434/v1 for host Ollama, http://host.docker.internal:8080/v1 for host llama.cpp, or the corresponding Unsloth port. On Linux, the container needs the host.docker.internal:host-gateway mapping. That mapping alone does not make a host loopback listener reachable on Linux. Use a host-network configuration supported by your Docker setup, or bind the model server to a Docker-reachable interface with authentication and firewall restrictions. Do not expose an unauthenticated server to your LAN. Test /v1/models from the same container or environment that sends OpenHands’ model requests before starting a task. Run the tool-cycle check. If you prefer automatic setup with OpenCode, rec handles it in one command.

​
Connect Flint

For the separate Nebius partner endpoint, use this profile. Open Settings → LLM → Advanced and create a profile. If a Provider Connection is selected, choose None so you can enter the URL and key. Keep the nebius/ prefix in Custom Model. Select Save Changes and make Flint (Nebius) active. If your version shows Settings → Agent, select an OpenHands agent and this LLM profile.

​
Try a task

Start a new conversation and ask:
Then run the file-editing and shell-check task to verify tool execution. If the app will not start, check that Docker is running. If your files are missing, stop OpenHands with Ctrl+C and run the launch command again from the project directory. For authentication or model errors, recheck the profile and start a new conversation. References: OpenHands LLM profiles and LiteLLM’s Nebius provider.