Models and providers

bitsrun does not come with a model. You pick one the first time you start it, and you can save as many as you like and switch between them in the middle of a session.

Set up a model

Setup opens on the first start. Open it again at any time with /connect inside the agent, or from your shell:

bitsrun code connect

The menu lists, in this order:

1. BitsRun (recommended). Use one of your private endpoints, or deploy any open model by its Ollama name or Hugging Face repo in one step. You sign in through the browser if you have not yet.
2. Model servers on this computer. bitsrun looks for Ollama, LM Studio, llama.cpp and vLLM at their usual addresses and lists the models they serve. No account needed.
3. Popular providers: Anthropic, OpenAI, Google, OpenRouter, Groq, Mistral, xAI, DeepSeek, Amazon Bedrock, Google Vertex and Azure.
4. Another provider. A search over many more providers from the public models.dev catalog.
5. Any OpenAI-compatible URL or any Anthropic-compatible URL, for a server or gateway of your own.

For a provider other than BitsRun, bitsrun asks for your API key, makes one test call, and saves the connection only if the call works.

Using a BitsRun endpoint

Choose BitsRun, then pick an endpoint you already have or Deploy a model (an Ollama name or a Hugging Face repo). When you deploy from here, bitsrun shows which GPU the model needs and what it costs per second before anything is created.

bitsrun then creates an API key that works only with that endpoint and keeps it in your system keychain. You can see and revoke it in the dashboard under API keys. To connect a specific endpoint directly:

bitsrun code connect --endpoint "Code helper"

bitsrun does not make a test call to a BitsRun endpoint, because that call would wake a GPU. Your first prompt shows Booting up while it starts instead.

Where your keys are kept

• In your system keychain: macOS Keychain, Windows Credential Manager, or the Secret Service (libsecret) on Linux. Keys are handed to the keychain on standard input, never as command arguments, so they do not show up in process lists.
• From an environment variable: if the provider's usual variable is already set (for example ANTHROPIC_API_KEY or OPENAI_API_KEY), bitsrun reads it each time and copies nothing.
• On a server without a keychain: bitsrun stores the key in a file only you can read, and tells you so during setup.

The list of connections (provider, model and address, never the key) is in code.json in the config folder.

Switch models

Type /model to pick another saved connection or set up a new one. The conversation continues with the new model. Choose Set up another model from the same list to add one.

Tool calls

The agent works by asking the model to call tools: read this file, run this command. A model that cannot make tool calls can only chat, and bitsrun tells you so when you pick it ("This model cannot make tool calls") and on the start screen.

• BitsRun endpoints report whether the model supports tool calls. For Ollama names it depends on the model's chat template; bitsrun tells you if it does not.
• Hosted providers generally support tool calls on their current models.
• Local servers depend on the model and on how the server was started.

Signing in with a subscription

bitsrun connects to providers with an API key. Signing in with a consumer subscription (such as ChatGPT Plus or GitHub Copilot) is not available yet, and Claude Pro and Max plans cannot be used by third-party tools under Anthropic's terms. For Claude, use an Anthropic API key, Amazon Bedrock or Google Vertex.

Licenses

bitsrun uses open source libraries to talk to providers. See their notices with:

bitsrun licenses