Three ways of working compared
Not everyone wants to work with AI the same way. One person prefers a web app; another spends the day in Claude Code or another AI tool and does not want to leave it. Plainwater AI supports three ways of working. The same data access rules apply in all three, because those rules sit on the tool layer, not in the model.
The three ways
| Fully online | With your coding AI | MCP only (Solo) | |
|---|---|---|---|
| Where you work | In the web app | In your own AI tool | In your own AI tool |
| Who directs | An agent in the platform | Your AI | Your AI |
| Who pays for the model | The platform (via credits) | Your subscription, plus the platform when you hand work over | Only your subscription |
| Subtasks by platform agents | Yes | Yes | No |
| Connectors, dashboards, datasets | Yes | Partly, via handover | Yes, directly |
| Data access rules | Apply | Apply | Apply |
1. Fully online
You work in the web app: conversations, work items, dashboards and monitors. The agents run in isolated containers in the platform. For Claude models they use the Claude Code harness; for other models a harness that works with any OpenAI-compatible model. A first responder answers simple questions straight away and passes heavier work on to an agent.
Suits: consultants and analysts who do not want to work in a terminal, and work that should continue while your laptop is closed.
2. With your coding AI
Your own AI tool connects to the platform through MCP and OAuth (see connecting your own AI tool). Your AI directs: it reads work items and messages, chooses sources and uses tools. When the work grows large, you hand it over: the platform puts an agent on it and you read the result back.
Suits: developers who already work in Claude Code, Codex or another MCP client and want to stay there.
3. MCP only (Solo)
Everything happens in your own AI tool, on your own subscription. The platform supplies the tools, the data access, the rules and the storage, and incurs no model costs of its own. Calling connectors and saving dashboards and datasets work directly too. See Solo.
Suits: anyone who already has a Claude Max or Pro subscription and wants no model costs on top.
What stays the same in all three
- Data access per source. The source's setting (metrics, encrypted values or open) applies to every tool, wherever the model runs.
- Pseudonymisation. Names and values come out of the tools as pseudonyms, also for your own AI.
- Permissions. Every call reads your permissions from the database again. You cannot do anything in your own AI that you are not allowed to do in the web app.
- Audit. Every tool call is recorded on the work item.
Local sessions are off by default per customer. A consumer subscription may train on content, and conversation messages can contain customer information. The customer decides whether to allow local sessions.
What ways 2 and 3 do not have
In a local session the model runs outside the platform. The platform therefore does not see those model calls and cannot take a policy decision per call. All oversight then sits on the tool layer: what your AI receives goes through the same gate as for a platform agent.
Example
A BI consultant spends the day in Claude Code on a data model (way 2). For an extensive performance analysis she hands the work over; a platform agent works it out overnight in a disposable copy. The next morning she reads the result in the web app (way 1). Meanwhile a colleague with a Claude subscription of their own builds a dashboard on the same source in Solo (way 3).