Set up AI (providers & keys)
Liza uses AI in three places: the AI assistant (chat/analysis in project context), the coding agents (desktop app), and scheduled AI tasks.
Two operating modes
| Liza AI (included) | Your own access (BYOK) | |
|---|---|---|
| Setup | none – it is simply there | store the provider's API key |
| Billing | the workspace's AI credit | directly with the provider |
| Available in | paid plans | all plans (Free: one provider) |
| Providers | operated by Liza, hosted in Europe | Claude, ChatGPT, Gemini, Perplexity, Mistral … |
In the AI chat you pick the provider via the pills above the input field. Only what is included in your plan, configured and allowed by your organization shows up.
Models of the included AI
With the included Liza AI you pick not only the provider but also the model – in the AI chat, in the settings, and in the AI status. All of them run through our access; you configure nothing and settle nothing directly.
| Model | In short |
|---|---|
| GLM-5.3 (default) | balanced, agentic, with caching – the standard choice |
| GLM-5.3 Flash | fast and cheap, for simple tasks |
| Qwen3.8 | the strongest model, priced accordingly |
| Qwen3.8 Flash | fast and cheap |
| Kimi K3 | very long contexts, the most expensive model |
All of them are operated in Europe. The models differ considerably in price – by up to a factor of 30 between the cheapest and the most expensive. Since they are paid from the same AI credit, picking a Flash model deliberately pays off for simple tasks.
Not every model is always available
The operator of your Liza environment can block individual models. A model that is not released does not appear in the picker – and a job using it is rejected with a note saying the model is currently not released. This has nothing to do with your credit.
Your own access: your own models
With your own key the provider decides which models exist – with Claude, for instance, Opus, Sonnet, and Haiku. That list comes live from the provider and can change without us releasing anything.
Personal: your own keys
Settings → Personal → AI settings – five tabs:
| Tab | Purpose |
|---|---|
| My AI keys | One key per provider – applies everywhere (assistant + coding agent) |
| General | Base settings of the coding agent (desktop only) – see Agents |
| MCP server | API keys & MCP for programmatic access (developers) |
| Usage | Your AI credit and consumption (only with included AI) |
| Memory | What the AI has remembered – see AI memory |
Next to it sit two entries of their own: AI providers (which AI you use) and AI agents and skills (see Agents).
Create keys with the respective provider:
| Provider | Create a key at |
|---|---|
| Claude (Anthropic) | console.anthropic.com/settings/keys |
| ChatGPT (OpenAI) | platform.openai.com/api-keys |
| Gemini (Google) | aistudio.google.com/app/apikey |
| Perplexity | perplexity.ai/settings/api |
| Mistral | console.mistral.ai/api-keys |
| Black Forest Labs (images) | dashboard.bfl.ai |
Precedence: your personal key takes precedence over the org key. On the desktop it is additionally stored locally encrypted (keychain) so the coding agent can use it without a server – one field, both stores.
Organization: central keys & policies (admin)
Settings → Organization → AI – four tabs:
| Tab | Purpose |
|---|---|
| Providers | Store providers' API keys centrally (encrypted). Assigned members then use AI without a key of their own. |
| Assignment | Decide per person which AI they may use – or release a key for everyone. |
| Usage | Org-wide consumption. Requires a separate admin/usage key from the provider (not the regular key). |
| Policies | Central lock: which AI providers are allowed in this organization at all. |
Policies (central lock)
Blocked providers disappear from the AI chat selection for all members and cannot be used with a personal key either. There is also a switch "Forbid personal API keys" – then only central org keys and the included Liza AI remain.
Local desktop agents
A coding agent running on the machine with its own provider login (e.g. your own Claude or ChatGPT account) bypasses Liza. The lock hides it in the client but cannot technically enforce it.
AI credit (included Liza AI only)
The included AI is paid from a credit balance in euros. Purchased credit does not expire.
- Personal credit – topped up per person.
- Company credit – a shared pot. Per person you set a limit for how much they may spend from it per billing period (empty/0 = no access, or explicitly "no limit").
- Topping up runs through Stripe's secure checkout; the credit is booked after successful payment.
This is managed under Settings → Organization → AI usage. In the AI chat an indicator shows the remaining credit; at 80% you get a warning, at 0 no new job starts.
Own keys ≠ credit
With your own key the Liza credit is irrelevant – you pay the provider directly. Credit notices only appear for the included AI.
Web vs. desktop
| Web (browser) | Desktop app | |
|---|---|---|
| AI assistant (included or own access) | ✅ | ✅ |
| Personal/org AI keys | ✅ | ✅ (+ local keychain) |
| Coding agents (files & commands) | ❌ | ✅ |
| Edit local files | ❌ | ✅ |
| Start/stop local dev servers | ❌ | ✅ |
The coding agents run in the desktop app only, because they edit local files. Scheduled AI tasks, on the other hand, can run managed on the server – no desktop needed, see Schedule.