Claude, GPT, Gemini or Mistral: one platform for every model
Meerwijk Care Group wants Claude for complex analysis today, and perhaps a European model for anything involving client data tomorrow. A year from now there will be another model that is better or cheaper. Tie your way of working to one vendor and you rebuild that choice every time.
In Plainwater AI a model is a choice per customer and per conversation. Rules, data and history stay where they are when the model changes.
Updated .
Which models does Plainwater AI work with?
Plainwater AI works with Claude, GPT, Gemini and Mistral, and with any model that has an OpenAI-compatible API. Each model has an explicit route to a provider.
| Model | How the platform reaches it |
|---|---|
| Claude (Opus, Sonnet, Haiku) | Through Azure AI Foundry |
| OpenAI GPT | Through Azure OpenAI, with EU data residency |
| Google Gemini | Through Gemini's OpenAI-compatible layer |
| Mistral | Directly; a European provider |
| Other models | Any model with an OpenAI-compatible API, as configuration |
Who decides which model is used?
The customer decides which models are allowed, and users choose within that set per conversation. With no preference you pick Auto and the platform chooses a suitable model.
For client records, Meerwijk only allows models with European processing. For the source with public figures Claude is fine too. That choice lives in the rules, and the gateway enforces it on every request.
A request to a model that is not allowed is refused before it reaches the provider.
What does it take to add a new model?
A model with an OpenAI-compatible API is configuration, not code. You add a route with an address, a way to authenticate and a name for the audit trail.
The new model immediately falls under the same gateway, pseudonymisation and audit as everything else. There is no side path around the rules.
Does every model work the same way as an agent?
Every model gets a harness that suits it, with the same tools and rules. Claude runs in the Claude Code harness, other models in opencode or Codex.
Ruben never sees this. He picks a model in the conversation and carries on; every tool call goes through the same gateway.
Can I use my own AI tool as well?
Yes, with any MCP-compatible AI tool, such as Claude Code, Codex, GitHub Copilot and Cursor. The model then runs on your own subscription and the platform provides the tools and the rules.
That is how Mila works from her editor. How it works exactly, and what does and does not pass through the platform, is covered under the three ways of working.
Where it stops
Not every model can do everything. A model with a non-standard tool-calling schema works less well in an agent that runs on tools, and you will see that in the results.
We only name models and tools we have set up or use ourselves. Other OpenAI-compatible models and other MCP clients should work, but test them first.
A vendor's newest models are not always available with European processing right away. If you want EU only, you may sometimes pick a generation older.
Frequently asked questions
Can I pick a different model per conversation?
Yes, within the models the customer allows. With no preference, choose Auto.
Does Plainwater AI work with GitHub Copilot or Cursor?
Over MCP, any MCP-compatible AI tool works, such as GitHub Copilot and Cursor. Claude Code and Codex are the clients the connection was built and tested with.
Can I allow European models only?
Yes. Per customer you decide which models are allowed, for example only Mistral and Azure OpenAI with EU data residency.
Do my rules still apply when I switch models?
Yes. Pseudonymisation, data access per source, budget and audit live in the gateway and apply to every model.