Skip to content

Set up a model provider

The AI Assistant needs a language model. Open the AI Assistant tab, press ⚙, choose one of three providers, fill it in, press Test connection, then Save.

AI Assistant settings

Provider You supply Billing
Claude subscription A Claude Pro or Max account, signed in through Claude Code on the computer running the app Your subscription
Anthropic API An API key from console.anthropic.com, or an organisation gateway's key and URL Per token, to that account
OpenAI-compatible A base URL, a model name and (if required) a key: OpenAI, a vLLM server, Ollama, CBORG or another gateway Depends on the server

Test connection sends one real request. A pass means the provider works end to end. On failure it shows the HTTP status and a hint:

Status Meaning
401 / 403 Wrong key or wrong key header
404 Wrong URL, or the endpoint speaks the other API format
400 Usually a wrong model name

Claude subscription

Claude subscription settings

This provider uses Claude Code, signed in with your Claude account. It is set up once on the computer that runs the app (on a cluster: on that login node).

  1. Install Claude Code.

    irm https://claude.ai/install.ps1 | iex
    
    curl -fsSL https://claude.ai/install.sh | bash
    
  2. Open a new terminal and sign in. A browser opens; sign in with your subscription's email. Add --sso for organisation single sign-on.

    claude auth login
    
  3. Check the login. The output must contain "loggedIn": true.

    claude auth status
    
  4. In the app, choose Claude subscription and press Re-check. The green line shows the path and version of Claude Code. If the app was already running when you installed Claude Code, restart the app first.

  5. If Re-check cannot find Claude Code, enter its path in Path to claude. On Windows it is usually %USERPROFILE%\.local\bin\claude.exe.
  6. Model is optional: sonnet, opus, haiku, or empty for your account's default.
  7. Press Test connection, then Save.

The app never sees your Claude credentials. A Pro or Max subscription does not include API keys; this route is how the subscription is used from the app.

Anthropic API

Anthropic API settings

  1. Choose Anthropic API.
  2. Paste the API key from console.anthropic.com.
  3. Pick a Model, for example claude-sonnet-5-5.
  4. Press Test connection, then Save.

Through an organisation gateway

If your organisation gives you a key for its own endpoint, find out from its documentation which format the endpoint speaks.

  • OpenAI format (paths end in /chat/completions, or the documentation says "OpenAI-compatible"): use the OpenAI-compatible provider below. This is the simpler route.
  • Anthropic format (paths end in /v1/messages, or the documentation says "Anthropic SDK" or "Claude API"): choose Anthropic API, paste the key and enter the model name the gateway lists. The endpoint and key header are not in the dialog. Set them in settings.json in the app's AI folder (see API keys for the folder), with the app closed:

    {
      "anthropic_base_url": "https://gateway.example.org/anthropic",
      "anthropic_auth": "bearer",
      "anthropic_headers": "X-Project: my-project"
    }
    
    • anthropic_base_url: the gateway URL. The base, or the URL ending in /v1 or /v1/messages, all work.
    • anthropic_auth: bearer sends Authorization: Bearer <key>; x-api-key (the default) sends x-api-key: <key>.
    • anthropic_headers: extra headers, one Name: value per line.

    Keep the other keys of the file as they are. Alternatively, set the environment variable ACE3P_ANTHROPIC_BASE_URL before starting the app.

OpenAI-compatible

OpenAI-compatible settings

  1. Choose OpenAI-compatible.
  2. Base URL: everything up to and including /v1.

    Server Base URL
    OpenAI https://api.openai.com/v1
    vLLM on a NERSC GPU node http://<node>:8000/v1 or the proxy http://127.0.0.1:8077/v1 (see vLLM on NERSC)
    Ollama on your computer http://localhost:11434/v1
    A gateway (for example CBORG) The URL from its documentation
  3. API key: required by hosted services, optional for local servers.

  4. Model: the model name exactly as the server lists it.
  5. Press Test connection, then Save.

The model must support tool calling. Without it, the assistant can only chat and cannot read files or plot.

API keys

By default, keys are kept in memory and forgotten when the app closes. Tick Save API keys on this computer to store them in the app's AI folder (%LOCALAPPDATA%\ACE3P-GUI\ai-agent on Windows, ~/.local/share/ACE3P-GUI/ai-agent on Linux). The page never receives a key back: settings show only whether a key is set and its last four characters.

When no key is entered, the environment variables ANTHROPIC_API_KEY and OPENAI_API_KEY are used. HTTPS_PROXY is honoured.

Permissions

Setting Default Effect
Ask before writing or editing files On Every write shows a diff to approve.
Ask before running commands On Every shell command needs approval.