A localization MCP server for AI coding agents
Agents translate. Humans approve. Your agent adds strings while it writes the feature, Glotra translates them into your project's languages, and a person approves the result.
Connect your client
One hosted endpoint, signed in with OAuth. Follow the guide for your client: Claude Code, Cursor or VS Code. In Claude.ai and Claude Desktop, go to Settings, Connectors, Add custom connector, and paste the URL.
https://api.glotra.yaxbi.com/mcp
What happens when your agent adds a string
- 1
It looks first. The agent calls
search_keysso it can reuse a key instead of creating a duplicate. - 2
It writes one language.
add_keytakes a key name and text in the project's base language. The project sets which languages exist, so the agent can't invent one. - 3
Glotra translates the rest. DeepL fills every target language in the same call. The result reports each language as base, provided, machine_translated, pending or failed.
- 4
A person reviews. The new translations appear under Needs review in the dashboard. An editor approves them, or corrects them, which also clears the flag.
- 5
Your apps get the text. Export a file per platform, or pull it at runtime with a project API key.
Tools
Only the tools your approved scopes allow are offered. Without write access, an agent never sees add_key or update_key.
| Tool | What it does |
|---|---|
| list_projects | The projects you can reach, with each one's base and target languages. |
| get_project | One project's languages, key count and your permissions. |
| search_keys | Search or list keys and their translations, or only those awaiting review. |
| get_key | A single key with every translation, plural forms included. |
| add_key | Create a key from base-language text; Glotra translates the rest. |
| update_key | Edit, rename or re-translate a key. |
| delete_key | Remove a key (needs the delete scope). |
| export_translations | The same JSON the SDK export returns. |
How human review works
Text written by machine translation or by an agent is flagged, per language and per plural form. Text a person types isn't. The base language isn't flagged either, since it is the source the translations come from.
- Approve one value from the key list, or everything in a project at once.
- Editing a flagged value clears its flag. Saving a key leaves the flags on any language you didn't change.
- When an agent replaces an approved value, it is flagged again.
- Viewers can see the flags but can't approve.
- Approvals go in the key audit log, and the translation history shows whether each change came from a person, an import, machine translation or an agent.
Review doesn't block delivery. Flagged text is included in exports and SDK responses, so clear the list before you release.
Permissions
- You approve the connection. The consent screen sets read, write and delete scopes. Delete is off by default, and you can limit a connection to chosen projects.
- Your role still applies. Your Glotra role is checked on every call, so a viewer stays read-only even after approving write access.
- Tokens stay separate. An MCP token doesn't work on the dashboard API, or the other way round.
- You can cut it off. Every connection is listed under Connected apps, and a revoked app loses access on its next request.
- Writes are attributed. The history and audit log record the person and the app, for example
you@company.com (via Cursor).
Questions
No. Glotra is a remote MCP server over Streamable HTTP, so your client connects to a URL. You sign in through your browser; there is no API key to copy into a config file.
Any MCP client that supports remote servers and OAuth. We publish setup guides for Claude Code, Cursor and VS Code with Copilot, and it also works from Claude.ai and Claude Desktop as a custom connector.
It can. Review flags text; it doesn't hold it back. Flagged translations are included in exports and SDK delivery, so a backlog never breaks a build. Clear the Needs review list before a release if you want every string checked.
No. No MCP tool can approve a translation. A signed-in person with editor access does that in the dashboard.
Machine translation runs on DeepL. A language it can't handle comes back as pending in the tool result, and the agent tells you, so someone can add that translation by hand.
No. MCP is included in every plan and in the 14-day free trial.