Skip to content

Providers and keys

Hotline Agent, the built-in agent, runs on a model from an AI service you connect under Settings → Providers. The same connections draw pictures, hear and speak on calls. Keys, and the OpenRouter and xAI sign-ins, are kept in your computer’s keychain (the macOS Keychain, Windows Credential Manager, or the Secret Service on Linux). ChatGPT and GitHub Copilot sign-ins are kept in files only your account can read, in Hotline’s data directory. None of them is ever written into a conversation, a teammate’s files or a log.

Add provider lists every service not yet connected:

Provider Connects with
Anthropic API key
OpenAI API key
ChatGPT Sign in with your ChatGPT subscription
OpenRouter Sign in, or API key
Google API key
xAI Sign in with SuperGrok or X Premium+, or API key
Z.AI, and Z.AI Coding Plan API key
Groq API key
DeepSeek API key
Mistral API key
GitHub Copilot Sign in
Ollama Local A server URL, no key
Ollama Cloud API key
OpenAI-compatible Any endpoint that speaks the OpenAI API: URL, optional key, your own model IDs, and whether it takes Responses or Chat Completions

An OpenAI-compatible server on the Responses API does not have to match OpenAI’s stream exactly. Servers with a looser stream, such as Meta’s Muse models at api.meta.ai, work from Hotline 0.36.2.

Hotline connects to Anthropic with an API key only. It does not offer a Claude subscription sign-in, and never asks for or holds a Claude login. To work on a Claude subscription, add a teammate that runs Claude Code: Claude Code signs itself in with your account, and Hotline runs the claude CLI you installed. Anthropic’s plans and terms decide what that use costs and whether your plan covers Claude Code run from another app, so check them.

Who you pay depends on how each teammate signs in. A coding tool uses its own login; Hotline Agent uses the connection you add here.

Integration Signs in with Billed to
Claude Code The claude CLI’s own login, or ANTHROPIC_API_KEY The Claude account claude is signed in to, or the key’s Anthropic API account
Codex The codex CLI’s own login, or OPENAI_API_KEY The ChatGPT account codex is signed in to, or the key’s OpenAI API account
Cursor cursor-agent’s login, your Cursor account Your Cursor account
Grok Build grok’s login, your Grok account Your Grok account
Gemini CLI gemini’s own sign-in The account or key it signed in with
opencode The providers you set up in opencode Each of those providers
Muse Code muse’s own login Your Muse Code account
Agents from the registry, such as Google Antigravity Their own sign-in Their own account; Antigravity names its Google plan beside the agent

Hotline holds no key or login for these and counts none of it against a budget. When a tool isn’t signed in, the failed turn’s card offers Sign in with the methods the tool itself offers.

Provider Signs in with Billed to Chat budget
Anthropic API key Anthropic API, per token Counted
OpenAI API key OpenAI API, per token Counted
ChatGPT Sign in Your ChatGPT subscription Not counted
GitHub Copilot Sign in Your Copilot subscription Not counted
xAI Sign in with SuperGrok or X Premium+ That subscription Not counted
xAI API key xAI API, per token Counted
OpenRouter Sign in, or API key Your OpenRouter credits, per token. The sign-in makes a key on your account. Key counted; sign-in not counted
Google, Groq, DeepSeek, Mistral, Z.AI API key That provider’s API, per token Counted
Z.AI Coding Plan API key Your Coding Plan Listed at no cost
Ollama Cloud API key Your Ollama account Counted at listed prices
Ollama Local Server URL Nothing: it runs on your machine Not counted
OpenAI-compatible URL, optional key Whoever runs the server Counted with a key

A model with no listed price on a key is counted at a deliberately high guard rate, so its budget runs down faster than the bill.

Open a connection to see it. The top shows how it is connected, with Change connection, and Sign out for a sign-in or Remove key for a key. An OpenAI-compatible connection has Edit connection and Remove connection instead, and Ollama Local has Disconnect. Under it, Models in the picker chooses All or Only some, each switched on or off and saved at once.

Hotline ships a catalogue of the models each provider serves, and where the provider can list its own models, Refresh reads them again, so a newly released model is usable before Hotline ships an update. A model by ID adds one the list is missing.

For Ollama Local, Hotline lists whatever the server has installed. For an OpenAI-compatible endpoint you supply the model IDs, and Hotline lists them from the server when it can.

GitHub Copilot lists the models your account is offered, and sends each model to the endpoint Copilot names for it. A sign-in that has no record of those endpoints fetches them on its first Copilot turn. If a Copilot model still refuses with a message about /chat/completions, press Refresh on the Copilot connection.

Settings → General → New teammates → Model is the model a new Hotline Agent teammate starts on. Change any teammate’s model, and its reasoning effort, from the picker beside its name at the head of its conversation.

Use for → Images picks what draws when a teammate uses generate_image, starting on Automatic. See Pictures.

Calls, below it, picks how calls run, what hears and speaks on them, and the model that answers calls to the desk. See Calls and dictation.

Budgets, under Calls, shows what paid use has cost today and this month, as three budgets:

Budget Counts
Chat Hotline Agent teammates on an API key, and the call assistant
Voice Paid speech: hearing and speaking on calls
Images Pictures drawn on an API key

Open one to see what it went on and to set a daily and a monthly limit. There is no limit until you set one; a limit of 0 turns that kind of paid use off. When the Chat budget is spent, a teammate’s turn stops with “Chat budget spent” before anything is sent.

Sign-ins to a subscription, local models, the desk’s own speech models and coding tools such as Claude Code are not counted: their cost is the subscription’s, or nothing.

Teammates that run Claude Code, Codex, Cursor, opencode, Gemini CLI, Grok Build or Muse Code sign in through that tool, not through Hotline. Hotline holds no key for them. See Teammates and drivers.