MCP Server for Translation — ConveyThis for AI Agents

Summarize this post with: 7 min read

We built an MCP server for translation, and it changes how you check on a multilingual website. Instead of clicking through dashboards, you can now ask your AI agent — Claude, Cursor, VS Code, and more — direct questions about your ConveyThis projects and get answers in seconds. This post explains what an MCP server is, what your agent can do with ConveyThis, how to connect in under a minute, and why we made it read-only on purpose.

What is an MCP server?

The Model Context Protocol (MCP) is an open standard that lets AI agents talk to external tools and data sources in a consistent way. Think of it as a universal adapter: instead of every AI app inventing its own integration for every service, a service exposes one MCP server, and any MCP-capable client can connect to it.

A ConveyThis MCP server for translation does exactly that for your localization data. It runs at a single endpoint — https://mcp.conveythis.com/mcp — and speaks Streamable HTTP. You authenticate with a personal API key, and from then on your agent can read your translation data through a small, well-defined set of tools. No scraping, no brittle screen automation, no copy-pasting CSVs into a chat window.

MCP has quickly become the common language between AI assistants and the tools they work with. The clients your team already uses — Claude Code, Cursor, VS Code, and others — speak it natively, which means there is no plugin to install on our side and no SDK to learn on yours. If your assistant can add a remote MCP server, it can talk to ConveyThis. That portability is the whole point of building on a standard rather than shipping yet another one-off integration.

The key word is read. The ConveyThis MCP server never changes anything. We’ll come back to why that matters.

What can your AI agent do with ConveyThis?

Once connected, your agent can answer questions that would normally mean opening the dashboard and clicking around. Under the hood, the server exposes seven read-only tools, and each maps to something you’d naturally ask:

Which of my sites is least translated? The agent lists your projects and reads each one’s details to compare coverage.
How far along is my Spanish translation? It reads translation statistics and tells you the percentage complete, by language.
Find every page where we translated “checkout.” It searches your translated strings and returns the matches.
What’s our approved German term for “dashboard”? It reads your glossary so brand terminology stays consistent.
What translation rules are active on my store? It reads the rules currently applied to a project.
How many words have I used this month? It reads your billing status and plan usage — without exposing any way to change your plan.

Because the agent has structured access rather than a screenshot, the answers are exact. You can chain these into real workflows: “Compare Spanish and French progress, then list the ten least-translated pages on the site that’s furthest behind.” That’s a question, not a project.

This is AI agent localization in the practical sense — not “let the AI translate everything,” but “let the AI reason over the localization work you’re already doing.”

A real localization workflow, in one conversation

Here is the kind of thing that used to take a dozen browser tabs. Suppose you are about to launch German and want to know if you are ready. You ask your agent: “Across all my projects, which has the lowest German coverage, and how many words are still untranslated there?” The agent calls the projects and stats tools, compares them, and answers with a ranked list — no spreadsheet required.

You follow up: “On that site, find every page where we used the word ‘checkout’ and check it against our glossary.” Now the agent chains the search and glossary tools together, surfaces the pages, and flags anywhere the translated term drifts from your approved wording. You close with “and how many words am I going to burn this month if I finish German?” — and it reads your billing usage to tell you, without ever touching your plan.

None of those steps is a new feature we built for the agent. They are the same projects, stats, translations, glossary, rules, and billing views you already have in the dashboard — just reachable in the place where you are already thinking and writing. The agent is doing the clicking and cross-referencing for you, and because every answer comes from structured data rather than a screenshot, you can trust the numbers.

How to connect in under a minute

Connecting is deliberately boring, which is the point.

Create a key. In your ConveyThis dashboard, open the MCP / AI Agents page and create an API key. It looks like ct_mcp_… and is shown only once — copy it before you close the dialog.
Add the server to your client. Point your client at https://mcp.conveythis.com/mcp and send your key as a bearer token. In Claude Code that’s a single command:
```
claude mcp add --transport http conveythis https://mcp.conveythis.com/mcp \
  --header "Authorization: Bearer ct_mcp_YOUR_KEY"
```
Cursor, VS Code, the MCP Inspector, Cline, and Windsurf all use the same endpoint and bearer header — only the config file format differs. We have copy-paste blocks for each one on the connect guide.
Ask. Try “How far along is my Spanish translation?” and watch the agent answer from live data.

That’s it. One key, a few lines of config, and your agent is connected — at no extra cost, because MCP access is included with your existing ConveyThis plan.

Is it secure? (read-only by design)

We made one decision early and held it: the ConveyThis MCP server is read-only. Agents can read your translation data, but they can never change, publish, delete, or bill anything. There is no write tool to misuse, no “translate everything” button an over-eager agent can press, and no path to your billing settings.

That’s not a limitation we’re apologizing for — it’s the trust model. No write access means no accidental edits, no surprise charges, and no agent gone rogue. Your translations stay exactly as you left them, no matter how you prompt the agent.

The rest of the model follows the same conservative line. Each key is a personal access token, scoped to read-only access across all six data areas. Keys are shown once at creation and stored only as a one-way hash on our side — we never keep the key itself, so even we cannot read it back to you. Each one carries an expiry you choose at creation, and every key is revocable from your dashboard at any time, with revocation taking effect immediately. If a key ever leaks, you revoke it and mint a new one in seconds, and the leaked key stops working on its very next request.

Because the agent connects with your key rather than your password, it never sees your login credentials, and it only ever acts within the scopes that key was granted. That is the quiet advantage of a read-only, key-based design over handing an assistant the keys to your whole account: the blast radius of a mistake — or a misbehaving prompt — is bounded by construction, not by good intentions.

One-click connectors for Claude Desktop and ChatGPT (over OAuth) are on the way. For now, any client that supports a Streamable HTTP MCP server with a bearer token is fully supported today.

Try it

If you already use ConveyThis, your MCP server for translation is waiting — there’s nothing new to buy. Create an API key, paste it into your favorite AI client using the connect guide, and ask your agent its first question about your translations.

New to ConveyThis? Start a free trial, translate your site, and connect your agent when you’re ready.

We Built an MCP Server for Translation — ConveyThis