Locize MCP Server: Translate from Claude, Cursor & VS Code

The Locize ecosystem already handles the mechanical parts of localization well. Keys get reported automatically via saveMissing, the CLI instruments your codebase, and your CI pipeline publishes on deploy.

For everyone building with AI assistants and coding agents, the Locize MCP server brings the rest into the conversation: finding what is missing before a release, translating with the project's glossary and memory, sending drafts to review, publishing, branching. You describe what you need; the assistant does it in your project, with your permissions.

Updated November 2026: Locize is now in Claude's connector directory, VS Code installs the server with one click, and the server has 28 tools (22 at launch), including context and review tools. See what changed since launch.

What is MCP?

Model Context Protocol is an open standard that lets AI assistants connect to external tools and services. Instead of copying and pasting context into a chat window, the AI can read from and write to your actual systems, with your explicit authorization.

Claude, Cursor, VS Code and a growing list of other tools support it natively.

Connecting in 60 Seconds

Claude (claude.ai, Desktop, mobile)

Locize is in Claude's connector directory:

1. Open Locize in the directory, or search for "Locize" under Customize > Connectors
2. Click Connect; your browser opens the Locize sign-in
3. Sign in and grant the requested permissions

A Personal Access Token is created in your profile and can be revoked at any time. To add the server by URL instead, use Add custom connector with https://mcp.locize.app.

Claude Code (CLI)

claude mcp add --transport http locize https://mcp.locize.app

Then run /mcp in your session to sign in.

VS Code (GitHub Copilot)

Install Locize in VS Code with one click, or add it to .vscode/mcp.json:

{
  "servers": {
    "locize": { "type": "http", "url": "https://mcp.locize.app" }
  }
}

Cursor and other clients

Add https://mcp.locize.app as a remote MCP server in your client's settings and follow the sign-in.

CI, scripts and scheduled agents

Create a Personal Access Token with the scopes the job needs and pass it as a header, for example in Claude Code:

claude mcp add --transport http locize https://mcp.locize.app --header "Authorization: Bearer $LOCIZE_PAT"

A nightly translation routine built this way is in do you still need a TMS when your coding agent translates the app?

What You Can Do

Once connected, you talk to your translation projects in plain language. Some things that work today:

Finding gaps before a release:

"Find all missing German translations in the checkout namespace of my main project."

The assistant compares the target language with the reference language, plural forms included, and returns the keys that need attention with their English values. Long lists come in pages, so the answer fits the conversation.

Reporting new keys from a codebase conversation:

"I just added these three strings to the Settings page. Report them to Locize as missing keys in the latest version."

The assistant calls report_missing_keys, which only adds new keys and never overwrites existing translations. With automatic translation on, the new keys are translated into every target language, using the style guide, glossary and translation memory as context.

Translating with the project's terminology:

"Translate the new checkout strings into French and Italian. Use the glossary, the style guide and the translation memory, and push them for review."

The assistant reads get_glossary, get_styleguide and search_translation_memory before drafting, then pushes its drafts with update_translations and quality: "AI", so the project's needs-review rule marks them for review (by default, every one). The context tools post shows each tool.

Working the review queue:

"Which German values are waiting for review in the onboarding namespace, and what do the quality checks flag?"

get_translations with needsReviewOnly returns the queue; with raw it adds the confidence score and its reason for scored AI translations, and with issues what the editor's quality checks flag per key. list_review_proposals shows pending proposals with who or what proposed them.

Publishing as part of a deploy conversation:

"We're deploying v2.3 in 10 minutes. Publish the production version of my project."

The assistant publishes the version and waits for the job to complete, reporting back when the CDN is updated.

Branch workflows:

"Create a branch from the latest version for the checkout redesign feature, then merge it back when I'm done."

Project management:

"Add Japanese to all versions of my project." "Rename the 'payment' namespace to 'checkout' across all languages." "What's the current translation coverage for each language in the production version?"

For one complete run with the real output, from the feature to the provenance export, see a coding agent localizes a feature, step by step.

Available Tools

The MCP server exposes 28 tools (22 at launch), organized around the translation workflow:

Every tool and its parameters are in the MCP documentation.

Authentication and Access Control

The MCP server supports two authentication methods.

OAuth 2.0 is used automatically by Claude, Cursor, VS Code and other MCP clients that support it. When you connect, you sign in with your Locize account (email and password, or your company's SSO) and choose which permissions to grant. A Personal Access Token is created in your profile for that authorization; you can see which client created it and revoke it with one click.

Personal Access Tokens can also be used directly for CI/CD pipelines, scripts, scheduled agents, or any client that doesn't support OAuth. Create one in your profile under Personal Access Tokens and pass it as a Bearer token. PATs carry explicit scopes (read, write, manage, admin), so you can grant the minimum each use needs.

On top of that:

• Your role still applies. A token never does more than your project role and your language, version or namespace restrictions allow (details).
• Destructive tools are marked. Tools that delete, overwrite or publish, and the AI review run, carry the MCP destructive hint, so clients that honour it ask for your confirmation before running them.
• Agent writes are labelled. Every write over MCP carries an agent mark with the token's name (for example "Claude (OAuth)"), so the history and the provenance export do not report an agent's text as a person's.
• No conversation is stored. Per tool call the server logs the tool name, your user id, the project id and a masked token; it does not store your prompts or the conversation (data and privacy).

A Note on Multi-tenant and Branch Projects

The MCP server handles the full complexity of your Locize setup. For tenant and branch projects, translations are merged with the parent project when reading, the same way the editor works. find_missing_translations understands tenant overrides and identifies what's genuinely missing versus what falls back to the parent.

What changed since launch

Where to go from here

• When coding agents translate your app: what breaks, what to keep: the pattern behind the tools.
• A coding agent localizes a feature, step by step: one real run over MCP.
• Do you still need a TMS when your coding agent translates the app?: keeping a scheduled routine and writing into Locize.
• MCP documentation: setup options, every tool, scopes and troubleshooting.

No Locize project yet? Create one, then connect. The setup really does take about 60 seconds.