Skip to main content

Use Multilocale from an AI client

Multilocale's hosted Model Context Protocol server lets an AI assistant work with the translation projects in your Multilocale organization. It can inspect projects and localized copy, audit missing or duplicate translations, export a locale dictionary, and apply an exact project or phrase change after approval.

The production endpoint is:

https://mcp.multilocale.com/mcp

The endpoint is remote and uses OAuth. There is no package to install, server to self-host, or API key to paste into a chat client. When the client opens the Multilocale sign-in page, authorize the organization whose projects you want to use.

Connect from Claude​

In Claude, open Settings → Connectors, add a custom connector, and use the production URL above. Claude redirects you to Multilocale to sign in and approve the requested read and write scopes.

The Multilocale connector and its companion Claude plugin are being prepared for their public directories. Until a public listing is approved, the custom connector URL is the source of truth.

Connect from ChatGPT​

For development and private testing:

  1. In ChatGPT, open Settings → Security and login and enable Developer mode.
  2. Open Plugins, choose the plus button, and create an MCP connection.
  3. Give it the name Multilocale and enter the production MCP URL.
  4. Complete the Multilocale OAuth flow and review the discovered tools.

Multilocale's public OpenAI plugin submission is a separate review and publish process. A working custom MCP connection does not mean a public directory listing has been approved.

Connect from Codex and editors​

Codex CLI supports the remote OAuth server directly:

codex mcp add multilocale --url https://mcp.multilocale.com/mcp
codex mcp login multilocale

VS Code can register the same endpoint:

code --add-mcp \
  '{"name":"multilocale","type":"http","url":"https://mcp.multilocale.com/mcp"}'

For Cursor, Windsurf, and clients that read an MCP configuration file, use:

{
  "mcpServers": {
    "multilocale": {
      "type": "http",
      "url": "https://mcp.multilocale.com/mcp"
    }
  }
}

A client that only supports local stdio servers can bridge to the remote endpoint:

npx mcp-remote https://mcp.multilocale.com/mcp

Only publish compatibility claims for clients you have actually tested.

Useful workflows​

The connector has seventeen tools, including two tools that render inline views in hosts that support MCP Apps.

Review projects and copy​

  • List projects and open a bounded project overview.
  • Search phrase keys and values, then inspect one exact key across locales.
  • Check which locales are missing keys relative to the default locale.
  • Find duplicate default-locale values that may be candidates for consolidation.
  • Export a bounded locale dictionary preview in API-paginated pages; use the CLI for a byte-exact or larger file export.
  • List an organization roster with only display name and role returned; email addresses and internal member/role IDs stay excluded.

Manage localization​

  • Create a project with a default locale and initial locale set.
  • Update the default locale, complete locale list, or translation context.
  • Add a key and generate its configured target-locale translations.
  • Add a locale and generate translations for source keys missing there.
  • Share every locale row of one key with additional projects.
  • Replace one exact translation or permanently delete a key's locale group.

The remote MCP does not read or write files on your computer. Use the Multilocale CLI when you need to import or download translation files in a repository.

Approval and cost boundaries​

Read tools need the multilocale:read OAuth scope. Seven write tools need multilocale:write.

Adding a phrase or locale can invoke machine translation and consume translation capacity. Updating a translation overwrites the previous value. Deleting a phrase removes every locale row for that key; when a row is shared, the deletion can affect other projects too. A client should show the exact project, key, locale, prior value, sharing set, and expected translation scope before asking you to approve one of these actions.

Changing a project's locales field is a complete replacement, not an incremental patch. Preserve every locale that you did not explicitly ask to remove.

Data and tenant boundary​

The OAuth account selects one Multilocale organization. Tool calls go through the Multilocale API, which rechecks project and phrase ownership; the MCP server does not query the database directly. Cross-organization project identifiers return not found.

Translation values are model-visible when you ask a connected AI client to inspect them. Treat product copy as potentially private and connect only a client whose data policy fits the projects you intend to use. Inline views are self-contained and never receive the OAuth token.

You can remove the connection from the client at any time. Review Multilocale's privacy policy and terms before connecting.

Troubleshooting​

  • Authorization error: reconnect Multilocale and complete OAuth; do not paste a REST API key into the client.
  • Project not found: call the project list first. Names are scoped to the organization selected during OAuth.
  • No phrase found: use the project name returned by the project list, and search for the key before using the exact phrase read.
  • A dictionary says hasMore: call the export again with the next skip value until hasMore is false.
  • An analysis says partial: report the phrase, locale, or offset cap and do not claim unscanned copy is complete or unique.
  • A locale write is larger than expected: run the missing-translation report for that locale first and review the number of keys to generate.