AI-ready frontends
Frontends that AI assistants can read and operate
An AI-ready frontend exposes its actions as typed tools that an assistant can call: filtering a table, filling a form, opening a record or building a chart. The assistant on your page can change the view and the content in response to what the user asks, using the same rules and permissions as the regular interface. This is one of our strongest areas.
What it looks like for your users
A user types “show me last week's late orders from the north warehouse.” The assistant applies the filters, sorts the table and highlights the rows. The user can see exactly what changed and undo it. Or they ask for a summary chart, and it appears on the page using your own chart component.
How we build it
In-app tools
The frontend registers typed actions (applyFilter, openRecord, fillForm, navigate) with schemas the model can call. Each tool runs through the same code path as a button click, so business rules are enforced once.
MCP servers
We expose your product's capabilities through the Model Context Protocol, so assistants like Claude and other MCP clients can work with it.
WebMCP
For browser-based agents, we register page tools using WebMCP, the proposed browser standard for this. Agents call defined actions instead of guessing where to click.
Generative UI from your design system
When the assistant builds a view, it can only use components from your library. Generated screens match your brand and stay accessible.
Streaming responses
Answers and generated views render as they arrive, with clear loading and cancel states.
A readable page for outside agents
Semantic HTML and a clean accessibility tree let external AI agents understand the page as well.
Guardrails
Giving an AI control over an interface needs limits. We build in:
- Permission checks per tool, matching the signed-in user's role.
- Confirmation before anything destructive or costly: deleting, sending, paying.
- An undo stack for AI-driven changes.
- An audit log of which tool was called, with what input, for whom.
- Handling of page and user content as data, so text inside a document can't instruct the assistant to take actions.
FAQ
Frequently asked questions
01Do we need to change our backend to add an AI assistant?
Usually not much. Most tools wrap actions your frontend already performs. We add a thin layer that describes them to the model.
02Which AI models can this work with?
Any model that supports tool calling. The tool definitions are model-independent, so you can switch providers later.
03Is WebMCP ready for production?
It's still an emerging standard. We build so that your in-app tools work today and can be exposed through WebMCP as browser support matures.
Next step
Start with a review of what you have
Send us your repository or a staging link. We'll come back with the three changes that would make the biggest difference to your release speed, performance and reliability.