Connect via MCP

The Langsys API can be used directly from AI assistants that support the Model Context Protocol (MCP). This lets AI tools browse, search, and call the Langsys API on your behalf.

The server address is https://mcp.langsys.dev. If you connected earlier with https://mcp.langsys.dev/mcp, that address keeps working.

Quick Setup

Claude Code

claude mcp add --transport http --scope user langsys https://mcp.langsys.dev

The first time your assistant calls a Langsys tool, it opens your browser to sign in and approve access — no token to copy or paste. See Authentication.

Which scope?

claude mcp add defaults to --scope local, which registers the server only in the project you ran it from. Since Langsys is usually something you want everywhere, we recommend --scope user:

Scope	Available in	Use it when
user (recommended)	every project on your machine, just you	you translate in more than one project
local (default)	the current project only, just you	you’re trying Langsys out in one repo
project	the current project, everyone on the team (checked into .mcp.json)	your whole team should get the server automatically

Scopes aren’t exclusive — a local or project entry of the same name takes precedence over your user one, so you can override the shared config per repo. Run claude mcp list to see what’s registered and from where.

Claude Desktop

Add the following to your claude_desktop_config.json:

{
  "mcpServers": {
    "langsys": {
      "type": "http",
      "url": "https://mcp.langsys.dev"
    }
  }
}

Cursor / Windsurf / Other MCP Clients

Use the HTTP transport URL: https://mcp.langsys.dev

Refer to your client’s documentation for where to add MCP server configurations.

What’s Exposed

Tools

The Langsys API endpoints an assistant may call are available as MCP tools, grouped by resource. Billing, payment methods, credits, coupons, account deletion, sessions and passkeys are done in the app, never by an assistant, and the token the sign-in mints cannot reach them.

• API Keys — create, list, update, delete
• Organizations — manage orgs, settings, users
• Projects — manage projects, settings, members
• Translatable Items — manage source content
• Translations — manage and machine-translate content
• And more — see the full API Reference for all available endpoints

Two tools are not API endpoints. account_status reads your account and tells your assistant the one next thing worth offering — finishing sign-up, a plan, languages, a first project, the SDK, automatic translation — or that everything is set. Your assistant calls it at the start of a conversation, so you just say hi. wait_for_plan waits, for up to 50 seconds per call, until a plan you are paying for in the browser is settled; your assistant calls it on its own after handing you the payment link, so the conversation carries on by itself once you have paid — you never need to come back and say so.

Resources

The server also exposes API documentation as readable MCP resources, so your AI assistant can browse endpoint details without making API calls:

• API Overview (langsys://api/overview) — base URL, authentication, and resource groups
• Endpoint docs by tag (langsys://api/endpoints/{tag}) — parameters, request bodies, and responses for each group
• OpenAPI Specification (langsys://api/openapi-spec) — the complete OpenAPI 3.0 spec as JSON
• LLM-friendly reference (langsys://api/llms-full) — full API reference formatted for LLM consumption
• Setup guide (langsys://setup/guide) — the full account-setup flow your assistant follows, step by step
• SDK guides (langsys://sdk/guides) — which SDK fits each framework, how to install it, where its guide is, and the API-key types

Prompts

• setup — checks your account and offers whatever fits where you are: finishing sign-up, choosing a plan, setting your organization’s languages, creating a first project, wiring the Langsys SDK into your app with an API key, or turning on automatic translation. Run it as /mcp__langsys__setup in Claude Code (the middle part is the name you gave the server) or pick it from the prompt menu in claude.ai. Your assistant makes the same offer on its own at the start of a conversation when something is missing, and stays quiet once everything is in place; the prompt lets you ask at any time. Payment never happens in chat: choosing a paid plan opens the payment step in the Langsys app.

Authentication

The Langsys MCP server is an OAuth 2.1 protected resource. You connect by signing in through your browser; there is nothing to copy or paste.

Add the server with no credentials, as in Quick Setup. The first time your assistant calls a tool, the server returns an authorization challenge and your MCP client opens a browser to sign in at app.langsys.dev and approve access. The client registers itself with the server (dynamic client registration), uses PKCE (S256), and receives its code on a loopback address (http://127.0.0.1, http://localhost or http://[::1], any port). The client stores the resulting token and reuses it — you won’t be prompted again until it expires, up to three months. This works with any MCP client that supports the MCP authorization flow (Claude Code, Claude Desktop, and others).

Troubleshooting

You keep being asked to sign in again

Claude Code stores the sign-in per server entry, not per URL. If a project also has its own langsys entry (added with --scope local or --scope project), that entry overrides your user-scope one in that project and needs its own sign-in, so signing in elsewhere doesn’t carry over. Check which entry a project is using:

claude mcp get langsys

If it shows Scope: Local config or Project config and you meant to use your user-scope entry, remove the override with claude mcp remove langsys -s local (or -s project).

When the server does ask you to sign in, its response says why: the token expired, it couldn’t be read, or the Langsys API rejected the session behind it. Signing out everywhere else from your Langsys account settings also ends your MCP sessions.

Tools disappear or return “unknown tool”

The server builds its tools from the Langsys API’s own documentation, and each tool’s name comes from its endpoint. When an endpoint changes, its tool gets a new name. A client that is still holding the old tool list will call a name that no longer exists and get “Unknown tool”. Reconnect the server (in Claude Code, /mcp → langsys → reconnect) to load the current list.