AITerm
Menu
Download free

AI Workflow

AI Providers and Models

In this handbookContents

AITerm brings your own AI. It never proxies your requests through an AITerm server: local providers run entirely on your Mac, and cloud requests go straight from your Mac to the provider you configured, authenticated with your own account. You choose the provider in Settings ▸ AI.

This page walks through every provider: what it is, how to set it up, where to get each value it asks for, and what to do when setup fails.

Built-in connection catalog

The connections below are offered out of the box. The list is generated from the app's own connection catalog, so it always matches the current build.

The built-in connection list below is generated from ConnectionCatalog.swift.

  • Local (Ollama) (Free)
  • OpenAI (Free)
  • Anthropic (Claude) (Free)
  • OpenRouter (Free)
  • Groq (Free)
  • DeepSeek (Free)
  • Mistral (Free)
  • xAI (Grok) (Free)
  • Gemini (Free)
  • Perplexity (Free)
  • Cerebras (Free)
  • Z AI (GLM) (Free)
  • MiniMax (Free)
  • Cohere (Free)
  • LM Studio (remote) (Free)
  • llama.cpp (remote) (Free)
  • MLX server (remote) (Free)
  • MLX (managed) (Pro)
  • Azure OpenAI (Pro)
  • Amazon Bedrock (Pro)
  • Claude Code (your subscription) (Pro)
  • Codex (your subscription) (Pro)
  • Custom OpenAI-compatible… (Free)

How your keys are stored

  • Cloud API keys are stored in the macOS Keychain, not in a plain config file.
  • Requests go directly from your Mac to the provider. AITerm does not see or relay your key.
  • Local providers (Ollama, managed MLX, or a custom server pointed at localhost) send nothing off the machine at all.
  • Before anything is sent to a cloud provider, secrets in your session context are redacted. See Privacy and Data Handling.

Local Ollama (Free)

Ollama runs open models directly on your Mac. It is the most private option: nothing leaves your computer.

Set it up:

  1. In setup, use the Run AI on this Mac door. If Ollama is already running, it says so (“Ollama is running here with … chat models — nothing leaves your Mac.”) and Use Ollama reuses it; if it has no chat model yet, the door says so and Pick a model opens the Model Manager. If AITerm’s engine is installed but stopped, Start local AI starts it. Or open Settings ▸ AI and start the local engine setup there.
  2. With no engine on your Mac, the door says “AITerm can download Ollama and a model sized for this Mac”, with that model’s size, and offers Download & set up. It does not download silently at launch — you confirm the size first.
  3. On first setup it pulls a default model so you can start immediately.

If a compatible Ollama engine is already running on your Mac, AITerm can reuse it instead of installing its own.

Managing models: press ⌘⇧M to open the Model Manager. From there you can browse installed models, pull a new model by name (for example llama3 or mistral), delete models, and switch the active one. A RAM-aware "Recommended for your Mac" list suggests models that fit your memory.

If it won't start: the Local AI Engine panel in Settings ▸ AI shows the status and offers Retry, reinstall, or update. Common causes it reports directly are a failed integrity check (use reinstall), not enough disk space (free space and retry), and being offline during a download (reconnect, or use a cloud key in the meantime).

BYOK cloud presets (Free)

"Bring your own key" presets cover OpenAI, Anthropic, and many OpenAI-compatible services (OpenRouter, Groq, DeepSeek, Mistral, xAI, Gemini, Perplexity, and more — see the generated catalog above). They are Free: you use your own provider account.

In setup, the Paste an OpenAI or Claude key door saves the key and checks it; every other service is set up in Settings ▸ AI.

Set one up:

  1. In Settings ▸ AI, pick the provider from the connection catalog.
  2. Paste your API key into the key field. It is saved to the Keychain.
  3. Pick a model. If the model list can't load, use the Retry button; you can still type a model id directly.

Where to get the key: from that provider's own account dashboard — for example your OpenAI account for an sk-... key, or your Anthropic console for an sk-ant-... key. AITerm does not issue these keys.

If a request fails to authenticate: the key is usually missing, mistyped, or expired. Re-open Settings ▸ AI and re-enter it. AITerm fails closed on a missing key rather than silently falling back to another provider.

Custom OpenAI-compatible server (Free)

Point AITerm at any server that speaks the OpenAI API — a local LM Studio or llama.cpp instance, a self-hosted gateway, or a private endpoint.

Open the Custom OpenAI-compatible server sheet from Settings ▸ AI and fill in:

  • Name — a label for the connection, e.g. My LM Studio.
  • Base URL — the server's OpenAI-compatible endpoint, e.g. http://localhost:1234/v1 (LM Studio) or http://localhost:8080/v1 (llama.cpp). This must be a valid URL or Save stays disabled.
  • Model id — optional: the default model to use when this server is selected.
  • API key — optional: leave blank if the server needs no auth; otherwise it is stored in its own per-server Keychain slot.

Each custom server keeps its own key slot, so servers never share credentials.

Azure OpenAI (Pro)

Azure OpenAI is a Pro provider. Open the Azure OpenAI setup sheet from Settings ▸ AI and fill in:

  • Name — a label for this connection (e.g. Azure East).
  • Resource — your Azure resource name, e.g. my-aoai. The sheet shows the resolved endpoint https://<resource>.openai.azure.com as you type.
  • Deployment — your deployment name, e.g. gpt-4o. In Azure this is the name you gave the deployed model, not the base model name.
  • API version — prefilled with a tested default (currently 2024-10-21); change it only if you need a specific version.
  • API key — your Azure OpenAI key.

The resource name, deployment name, and key all come from the Azure Portal (the Azure OpenAI resource's Keys and Endpoint page, and its Deployments tab) — AITerm can't supply these. Save stays disabled until the resource, deployment, and key are all present.

Amazon Bedrock (Pro)

Amazon Bedrock is a Pro provider. Open the Amazon Bedrock setup sheet from Settings ▸ AI:

  • Name — a label (defaults to the region if left blank).
  • Region — the AWS region, e.g. us-east-1. The sheet shows the resolved endpoint as you type.
  • Credentials — choose either an AWS profile or pasted access keys.
  • AWS profile mode — enter a profile name (default default). AITerm reads it from ~/.aws at request time and stores no secret itself. As the sheet notes: "Read from ~/.aws at request time — no secret is stored by AITerm. SSO / role / credential_process profiles resolve via the AWS CLI (v2.9+); if a profile's SSO session has expired, run aws sso login --profile <name>."
  • Static key mode — enter an Access Key ID (AKIA…), a Secret access key, and an optional Session token for temporary credentials. In this mode the keys are stored in the Keychain.
  • Model / profile id — the foundation-model id or cross-region inference-profile id, e.g. us.anthropic.claude-3-5-sonnet-20241022-v2:0. If AITerm can list your models you can pick from the dropdown; otherwise type the id directly (inference-profile ids aren't in the live list).

The region, credentials, and model ids come from your AWS account (IAM for keys, the Bedrock console for model access). If Save is disabled, the sheet tells you exactly which field is missing (for example, "Enter an AWS region to enable Save." or "Enter a model or inference-profile id to enable Save.").

If a profile uses SSO and the session has expired, run aws sso login --profile <name> in a terminal, then retry.

Managed MLX (Pro, Apple silicon)

Managed MLX is a Pro, on-device engine tuned for Apple silicon. Like Ollama it runs locally and sends nothing off the machine, but it uses a curated, AITerm-managed engine.

Managed MLX is set up from Settings ▸ AI (More… in setup).

Set it up:

  1. In Settings ▸ AI choose MLX. AITerm asks to confirm the download ("Set up MLX (managed)?") — it is size-aware and never downloads silently.
  2. After the engine installs, the Choose an MLX model picker opens. A "Recommended for your Mac" section suggests models sized to your RAM; a model larger than your available memory is flagged as possibly slow.
  3. Pick a recommended model, choose one from the live list, or type any mlx-community repo id, e.g. mlx-community/Qwen2.5-3B-Instruct-4bit.

Model repo ids come from Hugging Face (the mlx-community organization). The provider is only adopted once the engine is actually reachable. If setup fails, the panel shows the cause and offers Retry or a cloud-key off-ramp.

CLI subscription delegation (Pro)

If you already pay for Claude Code or Codex, AITerm can delegate to the official CLI already installed on your Mac and use your existing subscription — no separate API key. These are Pro connections ("Claude Code (your subscription)" and "Codex (your subscription)").

Set it up:

  1. Install the official CLI (Claude Code or Codex) so it's on your PATH. When setup finds it, it offers a subscription door — Use my Claude Code subscription, Use my Codex subscription, or Use my subscription when both are installed; without AITerm Pro it shows a line saying the CLI needs AITerm Pro instead. Settings ▸ AI lists both connections either way.
  2. Select the connection in Settings ▸ AI.
  3. Sign in when prompted — AITerm runs the CLI's own login in a terminal tab (claude /login for Claude Code, codex login for Codex). Authentication is handled entirely by the CLI; AITerm never sees a key.

AITerm can see that the CLI is installed, but not whether it is signed in until it asks. A connection check (Check in Settings ▸ AI; switching to the CLI runs one) says when the CLI isn't signed in, and a request that fails reports the CLI's own sign-in error — use Sign in / switch account in Settings ▸ AI and click Check.

If a request fails, the failure includes the CLI's own error output, which usually points at a sign-in or installation problem.

AITerm runs the CLI as a plain model, not as an agent. Every tool the CLI has of its own — running commands, reading files, web search, plugins and hooks — is switched off for the request, so a delegated request can only answer in text, and that text still goes through AITerm's own approval before anything runs on your Mac. This needs Claude Code 2.1.169 or Codex 0.153.3 or newer; an older CLI is refused, and the message tells you to update it (claude update or codex update).

Switching providers and Pro status

You can switch providers at any time in Settings ▸ AI. Switching to a cloud provider may show a one-time notice that this connection sends data off your Mac.

Pro providers (Azure, Bedrock, managed MLX, and the CLI-subscription connections) require an active Pro entitlement. If Pro lapses, AITerm falls back to a local provider rather than failing your session. See Free vs. Pro.

Model Manager

⌘⇧M opens the Model Manager for the active local provider — browse, pull, delete, and switch models without leaving AITerm.

Related pages: Settings Reference, Getting Started, Privacy and Data Handling.