On this page
Connect the way you already work Use a local model server Settings Center Check the connection Servers without a built-in runtime Related tutorials and researchConnect a model
Set up local models, institutional endpoints, and cloud connections.
Connect the way you already work
Run clio-coder configure, or choose Guided setup in the desktop alpha. First-time setup asks what you have: an API key, a subscription, a model server or gateway, or a setup file from your lab or team. When CLIO Coder finds a route first, such as a provider key in your environment, it lists that route for you to choose.
- Provide what the row asks for: paste a key, sign in to a subscription in the browser, or enter a server's address.
- For a model server or gateway, CLIO Coder checks reachability and reads the model list.
- CLIO Coder takes the provider's recommended model, or the only model a server lists. Otherwise you pick from the list. For a local server with no models, load one and choose Check again.
- CLIO Coder sends the model one short test request and saves when it answers. After a failure you can try again, change the key or address, or choose Save anyway. The first connection supplies the chat and worker defaults.
The test request shows that the model answers. It does not establish model quality or tool calling, and --no-verify skips it. Live, cached, and provider-catalog model lists are labeled separately. Choose a model that supports tool calling for repository work.
Already know the URL? Use clio-coder configure --quick for endpoint-first setup. A lab or team can publish one setup file: clio-coder configure --import <path-or-https-url> reads it, prints where prompts will go, and asks once before saving.
Use a local model server
Start LM Studio, Ollama, llama.cpp, vLLM, SGLang or Lemonade with a model that can call tools, then run clio-coder. It finds a server on its default port, such as Ollama on 11434, and offers it. On another address, run clio-coder configure and choose A model server or gateway. clio-coder targets --probe --tools --target <id> checks that the model really calls tools.
Settings Center
Run clio-coder configure --settings or choose Settings on the desktop. Both use these sections:
| Section | What to change |
|---|---|
| Connections | Providers, credentials, endpoints, and model lists |
| Chat | The connection and model answering you |
| Fleet | Worker models, profiles, and capacity |
| Context & Memory | Context limits, compaction, and memory |
| Permissions & Limits | Approvals, autonomy, and spending limits |
| Appearance | Output style and interface preferences |
| Integrations | Agents, libraries, plugins, and Git |
| Advanced | Diagnostics and the full settings editor |
Terminal /settings groups the same controls into its own areas: Recent & Pinned, Models, Chat and display, Workers, Context and memory, Safety and limits, Workspace, and Advanced.
Use connection default lets a route inherit its connection's model. In the terminal, save to the session, project, or global scope where offered. Configure edits global defaults; desktop saved settings apply to new conversations. Active conversation controls have their own scope.
Check the connection
Shell
clio-coder doctor
clio-coder targets --probeIf setup fails, check credentials, the endpoint address, and whether the local model is loaded. See troubleshooting for the next step.
Servers without a built-in runtime
Every connection names a runtime, which fixes how CLIO Coder talks to the server. clio-coder configure --list prints the runtime ids you can choose.
| Kind of server | Runtimes |
|---|---|
| Local model servers | lmstudio, ollama, llamacpp, vllm, sglang, lemonade |
| Cloud providers | openai, anthropic, google, mistral, groq, deepseek, openrouter, bedrock, and others |
| Any compatible endpoint | openai-compat, anthropic-compat, litellm |
| Another agent's login | claude-code, codex-cli, opencode-cli, pi-cli, and others; workers only |
A server that speaks the OpenAI or Anthropic API but has no runtime of its own usually works through openai-compat or anthropic-compat, or through a LiteLLM gateway. Local runtimes probe the server, report the loaded models, and price usage as free; the compatible runtimes make no local-server assumptions and never price as free.
When none of these fits, a developer can add a runtime as a plugin: a JavaScript file in the runtimes/ folder of CLIO Coder's config directory, or an npm package listed under integrations.runtimePlugins, which takes effect after a restart. A broken plugin is reported on stderr and never stops startup. The provider adapter cookbook describes the runtime descriptor a plugin exports.