MCP vs API: What Is the Difference?
An API is a set of endpoints that one program calls with code written for it. MCP (Model Context Protocol) is an open standard on top: an MCP server describes its tools so any MCP-capable AI app can find and call them the same way. MCP does not replace APIs. Most MCP servers call APIs underneath.
MCP vs API at a glance
People search mcp vs api, mcp server vs api, and the difference between mcp and api because the names get used as if they were rivals. They are different layers. An API is the service. MCP is a shared way for an AI app to discover and call tools, and those tools usually call an API.
As of November 3, 2026.
| API | MCP | |
|---|---|---|
| What it is | Endpoints one service exposes for programs | An open standard for AI apps to connect to tools and data |
| Who calls it | Code a developer wrote for that API | An AI app (the host) through an MCP client |
| How tools are found | You read the docs and hard-code calls | The server lists its tools, with descriptions and input schemas, at runtime |
| Message format | Varies (REST/JSON, GraphQL, gRPC and more) | JSON-RPC 2.0 |
| Transport | Usually HTTP | stdio (local) or Streamable HTTP (remote) |
| Auth | Whatever the API sets (keys, OAuth, signatures) | Optional in the spec; HTTP authorization is based on OAuth 2.1; many servers accept a bearer token |
| Typical use | App-to-app integration | Letting a model use tools and read context |
What an API is
An API (application programming interface) is a set of endpoints a service publishes so other programs can ask it for something. The caller has to know the URL, the method, the fields, and how to authenticate. Someone writes that code for that one API.
A weather app is the everyday case. The app does not measure the sky. It calls a weather API with a city name and an API key, and the API returns the forecast. If the weather company changes a field name, the app's code has to change too.
A trading example is an exchange API. Endpoints cover balances, orders, and fills. A bot calls them with an API key the exchange issued. Whatever that key is allowed to do, the bot can do. That is why exchanges tell users to turn withdrawal permission off on keys they hand to bots. The advice is general. The risk is the scope of the key, not the fact that an API exists.
What MCP is
MCP (Model Context Protocol) is an open standard. Anthropic introduced it on November 25, 2024, in the post Introducing the Model Context Protocol, and described it as open-sourcing the protocol that day. The docs at modelcontextprotocol.io call it "an open-source standard for connecting AI applications to external systems" and compare it to a "USB-C port for AI applications."
Three roles show up in the specification, version 2026-07-28. The host is the AI app. The client is the connector inside that app. The server offers tools. Messages are JSON-RPC 2.0, and the design is inspired by the Language Server Protocol. Servers offer tools (functions the model can call), resources (data and context), and prompts (templates). Transports are stdio on the same machine and Streamable HTTP on the network. Local stdio servers read credentials from the environment.
MCP server vs API
An MCP server is a program, often a thin one, that turns API actions into named tools with plain-language descriptions. An API endpoint such as POST /orders can become a tool named place_order with an input schema for symbol, side, and size. The model reads the description and decides when to call it. The server, not the model, holds the API key and makes the HTTP call.
That is the practical answer to mcp server vs api. The API is still there. The server is the adapter that lists a smaller, described set of actions. A trading MCP server for research and orders is covered on MCP for trading. This page stays on the general difference.
MCP vs REST API
REST is a style for building APIs: resources, HTTP methods, and usually JSON. MCP is a protocol for AI tool use. A REST API can sit behind an MCP server. They are different layers, which is the whole of mcp vs rest api. GraphQL or gRPC can sit behind a server the same way. MCP does not replace those styles.
MCP vs function calling
Function calling is how a model asks the app to run a function. The model returns a structured request: this name, these arguments. Your app then runs the function you defined and sends the result back. The function list is whatever you put in that app's prompt or tool config.
MCP is how the app gets those functions from outside servers. The host reads each server's tool list at runtime, then uses function calling so the model can pick one. They work together.
Is MCP going to replace APIs?
No. MCP servers need APIs, files, or other systems to do anything useful. The API still checks the request, moves the data, and enforces its own rules. What MCP replaces is custom glue code per AI app. Without it, every assistant needs its own integration for the same weather API or the same exchange. With it, one server can serve any host that speaks the protocol.
Does MCP use API keys?
It can. Authorization is optional in the spec, and HTTP authorization is based on OAuth 2.1 and bearer tokens (RFC 6750). Many remote servers accept a key in the Authorization header. That is the protocol. It does not mean Moby uses account sign-in. Moby's server accepts a bearer agent key.
The point that matters, once money is involved, is what the key can do. A key that can withdraw is a different risk from a key that can only request a quote. The protocol does not make a key safe. The server's tool list and the checks behind those tools do. More on that design is in AI agent security and AI agent guardrails.
Is MCP like an API gateway?
Partly. Both sit in front of other services. A gateway manages traffic, auth, and routing for APIs: rate limits, keys, and which backend answers. An MCP server describes tools for a model to choose from. Some teams put an MCP server behind or beside a gateway. mcp vs api gateway is that overlap, not a new product category.
How to turn an API into an MCP server
Four steps, and no more machinery than that. The official SDK list is on modelcontextprotocol.io.
- Pick the actions the model should have. Leave out anything it should not do.
- Define one tool per action, with a name, a plain description, and an input schema.
- Inside the tool, call the existing API.
- Keep secrets on the server. Expose only what the model should see.
A sketch of one tool, short enough to read in a minute:
name: place_order
description: Place a spot order
input: symbol, side, amount
call: POST /orders with the server-side key
return: the order idWhy the difference matters when the API can move money
A trading bot that holds a full exchange API key can do whatever that key allows. If the key can withdraw, a bug, a bad prompt, or a leaked file can move funds out. Exchanges and bot platforms tell users to turn off withdrawal permission on bot keys for that reason. This page does not name a vendor as an example of a leaked key. The lesson is the scope of the credential.
An MCP server can expose a narrower set of tools than the API under it, and it can enforce limits on the server, so the model never holds the broad credential. The spec itself warns: "Tools represent arbitrary code execution and must be treated with appropriate caution." It also says hosts must obtain explicit user consent before invoking any tool. Those lines are about care, not a promise that MCP is safe by itself. Safety comes from what the server exposes and what it checks.
Example: the Moby MCP server
Moby is one server that shows the pattern. It is not the subject of this page. The server is https://mcp.moby.win/mcp, Streamable HTTP, with the header Authorization: Bearer <MOBY_API_KEY>. The key starts with moby_agent_, is revocable, and is not a wallet private key. You create it in the Moby app at Agent tab, Connect. You can have up to 5 active keys.
Each user gets one agent, with its own wallet separate from your main in-app wallet. The agent cannot withdraw or send funds. The agent has no withdraw, transfer, or send tool. You set your agent's limits on max per trade, daily buy volume and market-cap range from the Moby app, and Moby checks them server-side on every buy your agent makes. Moby refuses agent buys on tokens it rates red. Rate limits are 6 trades a minute and 200 a day. Quotes last about 45 seconds. Moby Agents trade Solana tokens, paid in USDC, including tokenized stocks. Tell your agent your exits and it places real take-profit and stop-loss orders on Moby, including a one-call bracket. Buys have no automatic exits, so you name them in the instruction. A 0.75% swap fee (minimum $0.90 per swap), plus onchain fees.
A generic config block, with a placeholder instead of a real key:
{"mcpServers":{"moby":{"url":"https://mcp.moby.win/mcp","headers":{"Authorization":"Bearer <MOBY_API_KEY>"}}}}Host walkthroughs: Codex MCP setup, Claude Code MCP setup, and ChatGPT MCP. The tool list is on Moby MCP in the Agent Marketplace and in agents.md.
Frequently asked questions
What is the difference between MCP and an API?
An API is a set of endpoints that one program calls with code written for that API. MCP is an open standard on top: an MCP server describes its tools in a fixed format, so any MCP-capable AI app can discover and call them the same way. Most MCP servers call ordinary APIs underneath.
Is MCP going to replace APIs?
No. MCP servers usually wrap existing APIs. The API still does the work. MCP gives AI apps one common way to find and call that work, so each app does not need custom code for each API.
Is MCP an API?
In a loose sense, yes: it is a protocol with defined messages, built on JSON-RPC 2.0. But it is not a product API like a payments or exchange API. It is a standard for how AI apps and tool servers talk to each other.
What is the difference between MCP and function calling?
Function calling is a model feature: the model returns a structured request to run a function your app defined. MCP is the layer that lets an app get those functions from outside servers in a standard way. An MCP host usually uses function calling under the hood.
Does MCP use API keys?
It can. The MCP spec makes authorization optional and bases its HTTP authorization on OAuth 2.1, but many remote servers accept a bearer token or API key in the Authorization header. Moby's MCP server uses a revocable agent key sent that way.
Is MCP like an API gateway?
Partly. Both sit in front of other services. A gateway manages traffic, auth and routing for APIs. An MCP server describes tools for an AI model to choose from. Some teams put an MCP server behind or beside a gateway.
How do I turn an API into an MCP server?
Write a small server with an official MCP SDK, define one tool per API action with a name, description and input schema, and call the API inside each tool. Keep keys on the server side and expose only the actions the AI should have.
Try it with a Moby Agent
Set up a Moby Agent with its own wallet and your spending guardrails, then connect it to the AI you use.
You set your agent's limits on max per trade, daily buy volume and market-cap range from the Moby app, and Moby checks them server-side on every buy your agent makes.
Moby Agents trade Solana tokens, paid in USDC, including tokenized stocks.
Each user gets one agent, with its own wallet separate from your main in-app wallet. The agent cannot withdraw or send funds.
Connect it from Codex, Claude Code (CLI), Gemini CLI (with a paid Gemini API key or enterprise access) and Grok CLI, with the agent key. In ChatGPT, paste your personal Moby connector URL from the app. ChatGPT: works on Business, Enterprise and Edu workspaces that allow custom MCP servers (web). Any MCP client that can send an Authorization header can connect using the generic setup.
Tell your agent your exits and it places real take-profit and stop-loss orders on Moby, including a one-call bracket.
A 0.75% swap fee (minimum $0.90 per swap), plus onchain fees.
10% of the Moby fee back in $MOBY on every swap, 20% after your first swap, up to 30% as you climb tiers.
Start at Moby Agents. The product frame is agentic trading.
Trading crypto and tokenized stocks is risky and you can lose money. Nothing on this page is investment advice. Facts about other products are from their own sites as of November 3, 2026; check them for current details.