AI providers
Connect a hosted service or a server on your network and its models appear in the same picker as your local ones. Providers are optional and you add them yourself.
Add a provider
You can ask the agent to help add an AI provider; it opens the setup dialog so you enter the key there, rather than in chat. You can also open Settings → Connections and click Manage, then Add AI provider.
- Pick OpenAI, OpenRouter, or Custom compatible service for anything else that speaks the OpenAI API, such as a llama.cpp, vLLM or LM Studio server.
- Paste your API key. The dialog links to where you get one.
- Click Connect. LlamaBoss checks the key and loads the provider's model list.
For a local server with no key, choose No authentication under Advanced.
Choose models
Check the models you want in your picker. The search box filters long lists, and Refresh models reloads the catalog.
- Save keeps your checked models without changing the current chat.
- Save and use model also switches this chat to the highlighted model.
To add models later, open the provider again with Edit AI provider. It opens straight on the model list.
Test a model
Test selected model sends two tiny requests: a plain chat message, and one that asks the model to call a test tool. It reports whether each worked. This button tests Chat Completions connections; for a model routed through OpenAI Responses, test by sending a normal chat message and then a tool request with Agent Mode on. You can pick a thinking level for the test. The provider charges its normal rate for these requests.
Advanced connection settings
| Setting | Use |
|---|---|
| Base URL and chat path | Where requests go. The dialog shows the full resolved URL. The chat path is normally /v1/chat/completions. |
| Authentication | Bearer token, x-api-key header, or no authentication for a local server. |
| Saved connection and key name | Use a key you've already saved instead of pasting a new one. |
| Tool protocol | Native function calls for providers with OpenAI-style tools. XML fallback for models that follow LlamaBoss's text tool format more reliably. |
| Reasoning format | How thinking settings are sent. Auto detects it from the host. Set OpenAI, OpenRouter or Chat template yourself for a custom server. |
For services without a model catalog, Enter model IDs without connecting lets you type the IDs yourself. The key isn't checked in that case.
Some OpenAI models, such as gpt-5.6-luna and gpt-6-astra, use OpenAI's newer Responses API in LlamaBoss so tools and thinking can work together. When you use them through api.openai.com, LlamaBoss switches to it automatically. Leave Allow agent tools enabled. If a custom service uses an explicit Responses path, compatibility depends on that service.
Model options
Select a checked model to set:
- Display name shown in the model picker.
- Allow agent tools. Turn this off for models that reject tool definitions; they'll still chat normally.
- Generate images for image models. LlamaBoss asks for image output and saves the returned images as files.
In Advanced: model IDs the same options are written one model per line:
provider/model-id = Friendly name
google/gemini-image-model = Image model [image]
some-model = Chat only [no-tools]Service keys for Skills
Settings → Connections → Manage service keys stores keys for other services your Skills use, such as Gmail or RunPod. Each key is either saved in LlamaBoss or read from a Windows environment variable, and is passed to Skill scripts as an environment variable like RUNPOD_API_KEY. Stored service keys are supplied to supported script/helper runs; the persistent Python session does not receive those saved-key injections.
Keys pasted into LlamaBoss are saved as plain text in %LOCALAPPDATA%\LlamaBoss\secrets.json, protected by the Windows permissions on that folder. They are not encrypted. If you'd rather not keep a key on disk, use the environment-variable option.
What's sent
When you chat with a provider's model, the request can include your message, recent conversation, Project and Skill instructions, tool definitions and results, and any images or document text needed for the task. The provider handles that under its own terms. Switch back to a local model to process chat locally. Tools and scripts can still use the network and send data to the services they call. See Privacy & data.