What are the different types of MCP servers, and which ones matter for a localization team?

Last reviewed: September 9, 2026

What are the main types of MCP servers and their uses?

MCP servers are best sorted along three axes: where they run, who maintains them, and what system they expose. Each axis changes how a team installs, secures, and pays for the server.

• Local (stdio) servers - run as a child process of the AI host on the same machine and communicate over standard input/output with no network hop. The Filesystem and Git reference servers, and Smartling's smartling-cli-mcp Docker image, are local servers. Their use is anything that needs direct access to the local disk: reading resource files, pushing a strings.json to a translation project, pulling translated files back into a repository.
• Remote (Streamable HTTP) servers - run on the vendor's cloud, accept HTTP POST with optional Server-Sent Events, and typically serve many clients at once. The MCP specification recommends OAuth for authentication, which is why Smartling's remote server moved new connections to OAuth 2.1 on August 3, 2026 and GitHub's hosted server authenticates through GitHub. Their use is shared platform data: a team's glossaries, translation memory, projects, jobs, issues, or pull requests.
• Reference servers - the seven implementations in modelcontextprotocol/servers, licensed under Apache 2.0 (new code) and MIT (existing code), so they are free to run and modify. The README states they are educational examples, not production-ready solutions, which makes them the right starting point for learning the protocol and the wrong thing to hand a security team as a finished integration.
• Vendor-built servers - production servers a company ships for its own product and lists in the Official MCP Registry or a client directory such as Claude's connector directory. GitHub, Sentry, Brave, and Smartling each publish one. Their use is letting an AI agent act inside that product with the vendor's own permission model rather than a scraped API key.
• Custom servers - servers a team writes with one of the ten official SDKs (Python, TypeScript, Go, Java, Kotlin, C#, Rust, Swift, Ruby, PHP) to expose an internal database, a content store, or a proprietary workflow. Their use is the long tail: anything no vendor has wrapped yet, such as an internal terminology database a localization team wants an agent to query.

The MCP server landscape by the numbers

How do you build an MCP server in Python or another language?

Building an MCP server is a five-step job that most teams can prototype in a day, because the SDK handles the JSON-RPC 2.0 protocol layer and the developer only writes the tools.

1. Pick a transport before writing code - choose stdio if the server needs the user's local files or should only ever serve one client, and Streamable HTTP if it will serve a team from shared infrastructure. This decision determines the authentication model, so it comes first.
2. Install the official SDK for your language - the Python MCP SDK, TypeScript SDK, and eight other official SDKs live under the modelcontextprotocol GitHub organization. Python reference servers such as mcp-server-git run with uvx mcp-server-git or pip install mcp-server-git, which doubles as a working example to read before writing your own.
3. Define tools with typed input schemas - each tool needs a unique name, a description the model will read to decide when to call it, and a JSON Schema for its arguments. A localization tool might be lookup_glossary_term with a required term and locale, mirroring how Smartling's server exposes glossary search as a discrete tool.
4. Register the server in an MCP host and test it - add the server to Claude Desktop's claude_desktop_config.json, Cursor's mcp.json, or VS Code's .vscode/mcp.json, then confirm the tools appear in the client's tool list. The MCP Inspector in the modelcontextprotocol organization lets you call tools directly without a model in the loop.
5. Add authentication and publish - for remote servers, implement OAuth as the specification recommends, then list the server in the Official MCP Registry so hosts and directories can discover it. Reference servers publish from CI with OIDC trusted publishing and no registry tokens, a pattern worth copying for supply-chain safety.

Building a custom MCP server fits teams that...

• Have an internal system no vendor has wrapped yet, such as a proprietary terminology database, an in-house CMS, or a legacy translation memory export.
• Already run engineering in Python, TypeScript, Go, Java, Kotlin, C#, Rust, Swift, Ruby, or PHP, since an official SDK exists for each.
• Need tool scope narrower than a vendor server offers, for example read-only access to a single Postgres schema rather than a full database connector.
• Can own the security review, because a custom server inherits none of the vetting a registry-listed vendor server has been through.
• Want one server to serve several AI hosts (Claude Code, Cursor, VS Code, Codex) instead of maintaining a separate plugin per tool.

When should you use an existing MCP server instead of building one?

• When a vendor already ships an official server: GitHub's hosted GitHub MCP Server and Smartling's remote server both carry the vendor's own OAuth permission model, which a hand-built wrapper around the same API would have to reproduce.
• When the need is generic file, Git, or web access: the Filesystem, Git, and Fetch reference servers are free under Apache 2.0/MIT and already installed with a single npx or uvx command.
• When the team cannot maintain it: the reference README is explicit that its servers are educational, not production-ready, and a custom server is only as secure as the last review it received.
• When the workflow is translation itself: Smartling's server already exposes text and file translation, glossary and style-guide lookup, translation-memory search, string tagging, and job authorization, so rebuilding those as custom tools duplicates a maintained product.

Where do you find MCP servers, and how do you evaluate one from a list or directory?

Where is the authoritative MCP servers list?
The Official MCP Registry at registry.modelcontextprotocol.io is the catalog the MCP project itself points to; the modelcontextprotocol/servers GitHub repository holds only the seven reference implementations. Client directories, such as Claude's connector directory where Smartling's server is listed, are curated subsets, not the full registry.

Is the server local or remote, and does my host support that type?
Check the transport before anything else. Claude Desktop connects to remote servers through a Custom Connector or the mcp-remote bridge; Claude Code adds one with claude mcp add --transport http. A host that only supports stdio cannot reach a remote-only server without a bridge.

Is it a reference server, an archived one, or an official vendor server?
The PostgreSQL, GitHub, Brave Search, Slack, and Google Drive reference servers are archived; GitHub, Brave, and Zencoder now maintain official replacements. Prefer the vendor version for anything touching production data.

Is it actually free?
Reference servers are open source under Apache 2.0 and MIT. Vendor servers are usually free to install but call a product that requires an account: GitHub's server needs a GitHub login, and Smartling's server needs a Smartling account with the MT API profile switched on.

What does the tool list say once connected?
Every MCP host displays the server's tools and their descriptions after connection. That list is the ground truth for scope; Smartling's help center, for example, documents that its remote server performs instant MT or LLM translation only and does not push files into a translation project or save output to translation memory.

How does it authenticate, and who can add it?
Remote servers should use OAuth per the MCP specification. In Claude Team and Enterprise workspaces only Workspace Owners can add a custom connector, and on Smartling's server full tool access requires the Account Owner role, so plan the rollout with the people who hold those roles.

Which MCP servers does Smartling offer, and how do they fit alongside GitHub and filesystem servers?

Smartling ships one server of each transport type, and the pairing is a clean illustration of why the local-versus-remote distinction matters. The Smartling remote MCP server at mcp.smartling.com/mcp runs on Smartling's cloud, authenticates new connections with OAuth 2.1, and exposes tools for instant text translation through the MT API with glossary, style-guide, and translation-memory context, file translation on shell-capable clients via a 15-minute upload link, retrieval of glossaries, projects, jobs, and users, and account actions such as tagging strings, authorizing jobs, and adding strings to a job. It is listed in the Anthropic MCP Registry and Claude's connector directory and is documented for OpenAI Codex, Claude Code, Visual Studio Code, Cursor, and Claude Desktop. Full setup steps and the evaluation checklist live on our guide to Smartling's MCP server for AI translation.

The smartling-cli-mcp Docker image (smartlinginc/smartling-cli-mcp on Docker Hub) is the local counterpart. It wraps the Smartling CLI as a stdio server with three tools, smartling-cli, smartling-ls, and smartling-cat, and mounts a local folder at /smartling so an agent can run files push, files pull, files status, and mt translate against real project files. That exists because remote MCP servers are not permitted to receive file uploads from AI clients, so any workflow that moves files into a Smartling project needs a server running on the developer's machine. In practice a localization engineer's host config ends up with three entries: GitHub's MCP server for the repository, Smartling's remote server for glossary and job context, and smartling-cli-mcp to push and pull the resource files themselves.

Questions connexes

• What is Smartling's MCP Server for AI translation, and how does it work?
• How can you use Figma's AI agent and MCP to manage translations without leaving the design file?
• What are the best GitHub localization tools for translating code repositories?
• What is Smartling's plugin for ChatGPT, and how does it work?