Tuvi
Integrations

Connecting AI coding assistants to Tuvi over MCP

The fastest way to make an AI coding assistant useful for UI work is to stop feeding it screenshots and start feeding it structure. This guide wires an MCP-capable assistant to Tuvi in a few minutes.

What the integration gives you

Tuvi's built-in MCP server exposes a deliberately small toolset over streamable HTTP:

  • File listing — enumerate the design files the connected account can access;
  • Document structure — the node tree of a file: frames, components, instances, text;
  • Node properties — resolved layout values, fills, typography, component relations for any node;
  • Rendering — a PNG render of a node or frame, identical to what users see on the editor canvas;
  • Code generation — the Dev Mode generator, callable per node for the supported targets, from CSS and Tailwind to SwiftUI and Compose.

Together these turn "implement this screen" into a grounded task: the agent reads real values instead of estimating from pixels.

Setting it up

  1. In your MCP-capable client, add a server entry of type streamable HTTP pointing at your Tuvi MCP endpoint.
  2. Authenticate as the account whose files the agent should see — the agent inherits that account's permissions and nothing more.
  3. Verify the connection by asking the assistant to list available design files; you should see the same files that account sees in the web app.

That is the whole integration: no webhooks to host, no export pipelines to maintain, no per-tool SDKs.

Patterns that work well

  • Grounded component implementation. "Read the PricingCard component, render it, then implement it in our React codebase" — the agent pulls padding, tokens and typography from the file and adapts naming to your conventions.
  • Design-system linting. Walk a file's nodes and flag hard-coded values that should reference shared variables — a review that is tedious for humans and mechanical for an agent.
  • Documentation with real previews. Generate usage docs for a component set, embedding fresh PNG renders instead of stale screenshots.

Boundaries worth knowing

The MCP surface is read-and-generate: it does not mutate documents, so an over-eager agent cannot rewrite your design system. Rate limits apply per account, and access ends the moment you revoke the account's session. Treat the endpoint like any other credentialed integration: scope the account to the files the agent genuinely needs.

Structured design access is quickly becoming the difference between AI tooling that demos well and AI tooling that ships. With MCP in the box, Tuvi sits on the shipping side of that line.

Frequently asked questions

Which AI tools can connect to Tuvi?

Anything that speaks MCP over streamable HTTP — including Claude Code and other MCP-capable assistants and IDEs. The protocol is an open standard, so the list grows without Tuvi shipping per-tool integrations.

What permissions does a connected agent have?

Exactly the permissions of the account it connects as. The MCP surface is read-oriented (structure, properties, renders, codegen) and there is no bypass around file access rules.

What should an agent use renders for?

Visual grounding: after reading a node's structure, the agent can request a PNG render of that node — the same image users see in the editor — to verify its understanding before generating code.

See your designs in Tuvi

Import a .fig file on the free plan and edit together today.

See pricing