Running without an AI provider key
What the built-in development model can and cannot do.
On this page
The platform runs on one model, set in .env:
AI_PROVIDER=anthropic # anthropic, openai, google, openai_compatible or heuristic
AI_API_KEY=sk-ant-...
AI_MODEL=claude-sonnet-5-5
AI_BASE_URL= # required only for openai_compatible
Agents and the support assistant all use it. A workspace that adds its own model and key in its AI models settings uses that instead, for its own agents.
models
-
laguna Ready Default
poolside · poolside/laguna-s-2.1 · key ••••••••ABCD
-
Claude Sonnet 5 Not usable
Anthropic · claude-sonnet-5
A model with no key reads “Not usable” and agents fall back to the platform model. Nothing breaks; it just is not yours yet.
The one thing that catches people
On an install that caches its configuration, a change to .env is not seen until the cache is rebuilt:
php artisan config:clear # or config:cache
Nothing is copied into the database, so there is nothing to re-seed.
What a missing key looks like
Not a message about keys. The provider answers HTTP 403, the platform turns that into a failure, and what you see depends on where you were:
- in the support chat, "I cannot reach my assistant service at the moment"
- in an agent run, a step in the trace marked Error
If you see either, check AI_API_KEY and AI_MODEL first, or — for an agent on a workspace model — that model's key in the workspace settings.
The fallback, and what it is not
With AI_PROVIDER=heuristic (the default when nothing is set), agents run on heuristic-dev — a rule-based planner that matches your words to tool names.
It exercises the entire pipeline: planning, validation, policy, approval, execution, tracing and metering. Everything you see is real behaviour, not a simulation, which makes it genuinely useful for local development.
What it is not is a language model. It matches verbs and nouns, so "show me today's orders" works and an unusual phrasing may not. It refuses to run in production, deliberately. A keyword matcher quietly standing in for a model on real traffic is worse than an outage. Nothing about the answers would say so.
Using a different vendor
Change AI_PROVIDER, AI_API_KEY and AI_MODEL in .env and clear the config cache. Any OpenAI-compatible endpoint — OpenRouter, Together, Groq, DeepSeek, vLLM, Ollama — works as openai_compatible with its URL in AI_BASE_URL.
Sources and further reading
- Claude API overview — platform.claude.com
- OpenAI API quickstart — developers.openai.com
- Gemini API: getting an API key — aistudio.google.com
- Laravel: configuration caching — laravel.com