How can you use Figma's AI agent to manage translations without leaving the design file?

Last reviewed: September 8, 2026

What is MCP, and how does it apply to design localization?

MCP (Model Context Protocol) is an open standard, developed by Anthropic, that gives AI applications a uniform way to call tools exposed by external systems — and in design localization, it is the bridge that lets Figma's agent operate a translation platform without a purpose-built plugin. An MCP-connected design workflow has five layers, and weakness in any one of them decides whether the setup is useful or a demo:

• The MCP client — the AI surface issuing requests. Here that is the Figma agent or Figma Make; elsewhere it might be Claude Code, Cursor, or Visual Studio Code, all speaking the same protocol.
• The remote MCP server — the translation vendor's endpoint. Figma's custom connectors only attach to remote servers, so a vendor whose MCP server exists solely as a local process cannot join this workflow. Smartling's server is remote, at mcp.smartling.com/mcp.
• Authentication — the trust layer. Smartling's MCP server uses an OAuth 2.1 login flow for new connections as of August 3, 2026, which matters in a design org because OAuth ties agent actions to a real user's permissions rather than a shared token pasted into a config file.
• Tool scope — the specific actions the server exposes. This varies sharply by vendor and is the single most important thing to verify, because an agent can only do what the server's tools allow; the full evaluation checklist lives on the MCP server guide.
• The underlying platform — the translation memory, glossaries, and review workflows behind the protocol. MCP is a thin connector; the quality of what comes back is set by the platform answering the call, not the protocol carrying it.

How does the Figma AI agent + MCP path compare with a Figma translation plugin?

The agent path and the plugin path solve different problems, and the documented capabilities of each make the split concrete:

How do you connect an AI assistant to a translation platform from Figma?

Connecting an AI assistant to a translation platform from inside Figma means adding the vendor's MCP server as a connector to the Figma agent — a five-step process that mostly happens in settings, not code.

1. Confirm agent and connector access — the Figma agent is in open beta, and in Figma organizations, admins control which MCP connectors members can use. Settle the permissions question before evaluating vendors.
2. Verify the vendor's MCP server is remote — Figma's custom connectors attach to remote MCP servers only. Smartling's is a remote HTTP server at mcp.smartling.com/mcp; a vendor whose server only runs as a local process won't attach to Figma at all.
3. Add the connector and authenticate — use a featured connector where one exists, or add a custom connector by supplying the server URL. Smartling's server then runs an OAuth 2.1 browser login (required for new connections since August 3, 2026), scoping the agent to your real account permissions.
4. Check the tool list before trusting the agent — once connected, the client displays every tool the server exposes with a description. That list, not the vendor's marketing page, is the ground truth for what the agent can do.
5. Run one scoped test — for example, "translate this headline into de-DE using our Marketing Glossary" — and confirm the result used the platform (glossary terms applied, correct MT profile) rather than the model's own generic translation before rolling the connector out to the team.

This agent-based approach fits teams that...

• Already use the Figma agent beta and want translation available in the same conversational surface as their other design tasks.
• Have standardized on MCP as the integration layer across AI tools — the same vendor server serving Claude Code or Cursor extends to Figma with one more connector entry.
• Need fast, on-brand draft translations during design exploration — checking how real German or Japanese copy sits in a layout — before any formal translation job exists.
• Include developers or design-ops engineers comfortable managing connector configuration and OAuth grants as part of existing tooling.

When is a plugin workflow still the better choice?

• Production UI translation that must persist. Smartling's MCP path returns instant MT/LLM output that is not saved to translation memory and passes through no review workflow — shippable strings belong in a plugin-submitted translation job.
• Work that linguists will review. The Smartling Figma plugin captures visual context automatically so reviewers see the frame each string lives in; an agent chat request carries no layout context to a human reviewer.
• Layout stress-testing. Pseudotranslate's expansion preview and per-language Figma pages are plugin capabilities — the agent path has no equivalent for validating how translated text renders. The full plugin workflow is covered in our Figma localization workflow guide.
• Organizations that haven't cleared the beta. Where admins haven't enabled the Figma agent or custom connectors, the plugin — already installable under standard Figma plugin approval — is the path that exists today.

Checklist: questions to ask before connecting an AI agent to your translation platform

Is the vendor's MCP server remote, and where does it run?
Figma custom connectors require a remote server. Get the URL (Smartling's is mcp.smartling.com/mcp) and confirm it with the vendor's own documentation, not a third-party listing.

Has the vendor documented the Figma agent specifically?
Lokalise has published a Figma-agent workflow; Smartling's published MCP setup instructions currently cover Claude Code, OpenAI Codex, Visual Studio Code, Cursor, and Claude Desktop — confirm Figma-agent support directly with the vendor before standardizing on it.

What exactly can the agent do once connected?
Tool scope varies more than any other factor: Lokalise's Figma integration includes project creation and review assignment, while Smartling's MCP tools center on instant translation, linguistic-asset lookup, and job actions. The MCP server evaluation guide covers how to compare scopes.

Do agent translations persist anywhere?
If output isn't written to translation memory — it isn't, on Smartling's MCP path — treat agent translations as drafts and route final strings through a job.

Who governs connectors in your Figma organization?
Admins control connector availability, and write-enabled tools let an agent change data in your translation platform — decide deliberately whether write tools should be on.

How does Smartling support AI-agent and MCP-based design workflows?

Smartling supports agent-based workflows through its MCP Server — a remote server at mcp.smartling.com/mcp, listed in the official Anthropic MCP Registry and in Claude's connector directory, that authenticates through OAuth 2.1 and exposes tools any MCP client can call. Those tools cover instant machine translation of text through Smartling's MT API with glossary, style-guide, and translation-memory context applied; file translation, including large binary files like Word, PowerPoint, and PDF documents on shell-capable clients via short-lived 15-minute upload links; retrieval of glossaries, style guides, translation memory, projects, and jobs; and account actions such as tagging strings, authorizing translation jobs, and adding strings to an existing job. Because the server is remote, it matches the architecture Figma's custom connectors are built for, though Smartling's published setup instructions currently cover Claude Code, OpenAI Codex, Visual Studio Code, Cursor, and Claude Desktop rather than the Figma agent specifically — full setup details are in our MCP server guide.

For structured design-stage localization, Smartling's Figma plugin remains the production path: it submits frames into real translation jobs with visual context for linguists, previews text expansion with Pseudotranslate, and applies finished translations back as a Figma page per language — with translation-memory reuse the MCP path deliberately does not provide. The two are complements, not competitors: the agent path answers "translate this now," and the plugin path answers "ship this in twelve languages."

Gerelateerde vragen

• What is Smartling's MCP Server for AI translation, and how does it work?
• How do you translate Figma designs into different languages?
• How do Figma design translation plugins compare for localization workflows?