Skip to main content

Connect an AI Provider

Goal​

Point the platform at an AI model, once, for the whole organisation — so the AI button in every designer generates a design from a sentence, and the assistant on the home screen can answer questions and build things for you.

Prerequisites​

  • You are signed in to the Admin Control Center as a platform administrator.
  • You have an account with an AI provider and an API key from it — or a model running on a server of your own (Ollama, LM Studio) that needs no key.

Steps​

  1. Open Settings — the gear in the top bar, on any screen — and choose the AI tab. If nothing has been set up yet, the tab opens on Anthropic with the model claude-opus-5 already filled in — so connecting to Claude is steps 4 and 5 only.
  2. Pick the Provider:
    • Anthropic for Claude. There is one endpoint, so no root is asked for, and the recommended model is filled in for you.
    • OpenAI-compatible for OpenAI itself, OpenRouter, Ollama, LM Studio and anything else that speaks the same /chat/completions contract. Then fill in the API root — the root, not the full endpoint: https://api.openai.com/v1 (filled in for you), https://openrouter.ai/api/v1 or http://localhost:11434/v1.
  3. Check the Model — type it exactly as the provider spells it: claude-opus-5, gpt-4o, llama3.1. Switching provider fills in that provider's default model, but never replaces one you typed yourself. Leaving it blank turns AI off for your organisation.
  4. Paste the API key. Leave it blank for a local server that needs none. For Claude, see Getting an API key from Anthropic below.
  5. Save, then Test connection. The test makes one small call through what you saved and shows the model's own reply, so a green result means the provider actually answered.

Getting an API key from Anthropic (Claude)​

An API key is not the same as a claude.ai subscription. Signing in to claude.ai gives a person a chat; the platform needs a key from the Claude Console, which is a separate, pay-as-you-go developer account that bills the organisation for what its users generate.

  1. Go to platform.claude.com and sign in, or create an account. If your company already has a Console organisation, ask its admin to invite you to it rather than creating a second one — the key must belong to the organisation that pays.
  2. Add billing under Settings → Billing. A new account has to have a payment method or prepaid credits before its keys will answer; a key without credit fails the connection test with a billing error, not an authentication one.
  3. Open Settings → API keys (platform.claude.com/settings/keys) and press Create key. Give it a name that says where it is used — Parsware platform — choose an expiry, and set the Linked account: for a shared platform, prefer a service account key over a personal one, because a personal key stops working the day its owner leaves the organisation.
  4. Copy the key straight away. It starts with sk-ant- and the Console shows the whole thing exactly once, at creation; if it is lost, create another rather than looking for it.
  5. Back in the platform, paste it into API key, Save, and Test connection.

Keep the key in one place. The platform stores it on the server and never sends it back to a browser; do not also keep it in a document or a chat. If it ever leaks, revoke it on the same Console page and paste a new one here — nothing else needs to change.

What happens behind the scenes​

The key never comes back to the browser. After saving, the tab shows only that a key is stored and its last four characters. An empty key box on a later save means keep the one that is stored; to remove it, use Remove the stored key and save.

The settings are your organisation's, not each person's. Everyone in the organisation generates against the same model, and the cost lands on the same account. Before this tab existed the provider was set by whoever installed the platform, in its configuration; those values are still the fallback and are shown here as a starting point until you save your own — the line at the top of the tab says which is in effect.

The two limits are there for runaway replies. Longest reply caps one answer in tokens; Give up after is how long one call may take before the platform stops waiting. The defaults suit generating a process or a report; raise them only if a test says a reply was cut short.

Where it is used​

  • The AI button in the process, report, dashboard and Studio designers — describe what you want and a design is generated, checked against the design's own schema, and loaded onto the canvas.
  • The assistant on the home screen, which can read what is in the environment and, when you allow it, build tables, apps and whole solutions.