What is MCP? Connect MCP servers to your AI agents

What is MCP?

MCP defines how an AI application finds the tools a server offers, calls them and gets results back. Anthropic released it in November 2024 and in December 2025 donated it to the Agentic AI Foundation, part of the Linux Foundation. The protocol is open source and its specification is developed in public. Claude, ChatGPT, Cursor, VS Code and many agent frameworks speak it.

Before MCP, every AI app had to write its own integration for every service, and every service had to support every AI app. MCP turns that into one integration on each side: a service writes one server, and every MCP client can use it.

What is an MCP server?

An MCP server is a program that offers an AI application three kinds of things: tools it can call, such as create_issue; resources it can read, such as files or records; and prompts, reusable instructions for a task. The client asks the server what it has, shows the tool names and descriptions to the model, and runs a tool when the model picks one. Messages are JSON-RPC.

A local server is a program on the same machine as the agent, talking over standard input and output (stdio). A remote server lives on the web and is reached over HTTP, using the Streamable HTTP transport or the older SSE one. Remote servers that need an account use OAuth, so you sign in with the app in your browser and the agent never sees your password.

Most MCP servers sit in front of an API the app already has. The server doesn't replace the API; it describes a useful slice of it in a way a model can use.

MCP vs API: what's the difference?

An API is how programs talk to a service. MCP is how an AI application finds and uses tools, and the tools usually call an API underneath. You don't choose one instead of the other; you choose whether your agent talks to the API through an MCP server, or calls it directly through something like an API connector.

Are MCP servers free? What MCP costs

MCP itself costs nothing: the specification and official SDKs are open source, and there is no license or per-call fee for using the protocol. What you pay depends on who runs the server and what it calls.

Gateway prices from Zapier, Composio and Pipedream, checked November 2026. In Ork, MCP servers are included in every plan, on Ork Cloud and self-hosted.

• The hidden cost is usually tokens, not fees. A server with 40 tools can add thousands of input tokens to every model call. Ork counts tools by source, not per server: the MCP servers and OpenAPI and GraphQL connectors installed from Integrations share one count, so a few small servers and one large API can add up together. Up to 20 tools, a bot gets them all directly; past that, it searches for the tools it needs and loads only those.
• A gateway that covers thousands of apps costs money and sees your data in transit. If the app publishes its own server, connecting to it directly is cheaper and has one fewer company in the middle.

Good MCP servers to start with

The best MCP server for an app is usually the one the app's own team publishes, because they keep it in step with the API. Many well-known apps now host one with OAuth sign-in: GitHub, Linear, Notion, Sentry, Stripe, Atlassian, Asana, HubSpot, Cloudflare, Supabase and Vercel among them. Browse MCP servers in Ork searches integrations.sh for an app's server so you don't have to hunt for the URL.

Ork bots already have their own computer with a browser and a terminal, so the local servers popular in coding tools, for files, browsing or running commands, add little. Spend MCP servers on the apps your bots need to work in.

What bots can do with MCP servers in Ork

• Add a remote server by URL, or find an app's server with Browse MCP servers, which searches integrations.sh. Ork speaks Streamable HTTP and the older SSE transport.
• Sign in with OAuth. Ork registers itself with the provider when it allows that, stores the grant encrypted and refreshes it. Servers that take a bearer token or a custom header work too, and the credential is never sent to the model.
• Choose which bots use a server. A bot asks in the chat before its first use; choose Approve, or Authorize if you need to sign in.
• See when a server breaks. Test connection shows whether it works and why not. A bot whose server is down is told so instead of quietly losing its tools.
• Follow your approval settings. In Ask mode a bot asks before a call that changes something. Tools the server marks as read-only usually run without asking.
• Install a plugin from GitHub that bundles remote MCP servers with skills that teach bots to use them. Anyone can build one and submit it.

Zapier MCP and other hosted MCP gateways

Zapier MCP, Arcade, Merge Agent Handler and similar services put many apps behind one MCP endpoint. Ork has no built-in support for them, but any of them is just a remote MCP server to Ork, as long as it speaks Streamable HTTP or SSE and signs in with OAuth, a bearer token or a single custom header. Zapier MCP uses Streamable HTTP and OAuth, so you add its URL and sign in.

Two things to know. A gateway acts as whoever signed in to it, so every bot that uses it acts as that person's accounts. And it bills by its own meter: each successful Zapier MCP call spends 2 tasks from your Zapier plan.

When to choose MCP servers

Choose an MCP server when the app already publishes one. The app's own team decides what the tools do, there is no platform in the middle, and there is no fee beyond the app's own plan. If you host Ork yourself, MCP is also how you reach an internal tool or run a server on your own machine.

Composio and Pipedream are managed platforms with ready-made tools for many apps, which helps when an app has no good MCP server. If an app only has a REST or GraphQL API, API connectors turn it into tools. The comparison puts all four side by side.

Limits and security

• Remote servers must be at a public HTTPS address. Localhost and private networks are refused unless the operator sets MCP_ALLOW_PRIVATE_ENDPOINT=true on the API and worker, which allows them for every user. Cloud metadata addresses stay blocked.
• Stdio servers are off unless the API and worker have MCP_STDIO_ENABLED=true and the command is in MCP_STDIO_ALLOWED_COMMANDS. They run with the server's own access, so only admins, the verified accounts in ORKBOT_ADMIN_EMAILS, can add one. Bots never do.
• npx packages are installed once without running install scripts. Other runners such as uvx are refused; install those servers on the host and allowlist the binary.
• Plugins can only include remote servers, and Ork never runs code from a plugin's repository.
• Very large tool results are cut to fit the bot's context, unless you turn on Keep tool results whole for that server.
• A tool can return text anyone wrote, such as an email. Keep Ask or Auto approvals on for bots whose servers can make changes.

How to set it up

Anyone with an account can add a remote server, on Ork Cloud or a self-hosted server. Stdio servers are for admins of a self-hosted Ork. See Connect apps and the tools and integrations guide.

1. Open Integrations from Connect apps at the bottom of the sidebar and choose +.
2. Choose Browse MCP servers, search for the app and choose Connect. To enter a URL yourself, use Manage MCP servers → Add server.
3. Choose Connect OAuth and sign in if the server asks. If the provider only accepts a client you create in its console, open Advanced, add the Redirect URI shown there to that client, then enter its client ID, and its secret for a confidential client.
4. Ask a bot to use the server and approve its first use, or choose Use in this bot.
5. For a stdio server, an admin sets MCP_STDIO_ENABLED and MCP_STDIO_ALLOWED_COMMANDS on the API and worker, then adds it under Add server → STDIO.