6 min read
Also: MCP, MCP server, MCP client, Anthropic MCP, modelcontextprotocol, MCP protocol
Model Context Protocol (MCP)
Quick answer
The Model Context Protocol (MCP) is an open standard, released by Anthropic in November 2024 and now run under the Linux Foundation's Agentic AI Foundation, that lets AI apps call tools and read data through MCP servers over stdio or Streamable HTTP. To give a browser agent your own MCP tools, see Autonoly's AI agent.
The Model Context Protocol (MCP) is an open standard, created by Anthropic in 2024 and now governed by the Linux Foundation's Agentic AI Foundation, that lets AI applications discover and call external tools and data through a common JSON-RPC interface.
Last verified
What people do with Model Context Protocol (MCP)
- Connect custom MCP servers to an AI agent
Use your MCP tools alongside a cloud browser, terminal and integrations.
- Drive a browser from a model with Playwright MCP
Microsoft's Playwright MCP server and how browser agents use it.
- Understand OAuth for remote MCP servers
MCP's HTTP authorization is built on OAuth 2.x.
- Compare MCP with a classic API integration
When a tool server beats a direct API call, and when it does not.
- Choose an AI agent platform
What to look for in agents that use tools on your behalf.
Key facts
| Created by | Anthropic, open-sourced 25 November 2024 |
|---|---|
| Governed by | Agentic AI Foundation (Linux Foundation) since December 2025 |
| Message format | JSON-RPC 2.0 |
| Standard transports | stdio (local) and Streamable HTTP (remote) |
| Latest spec revision | 2026-07-28 (stateless; no initialize handshake) |
| Server primitives | Tools, resources and prompts |
What is the Model Context Protocol?
The Model Context Protocol (MCP) is an open standard for connecting AI applications to external tools and data. Instead of every AI app writing a custom integration for every service, a service is wrapped once as an MCP server, and any MCP-capable client (Claude, ChatGPT, an IDE such as Cursor or VS Code, or an agent platform) can discover and use what it offers. It is often described as "USB-C for AI": one plug shape, many devices.
Anthropic released MCP as an open-source project on 25 November 2024. Adoption spread quickly: OpenAI announced support in March 2025, and Google, Microsoft and most agent frameworks followed. On 9 December 2025 Anthropic donated MCP to the Agentic AI Foundation (AAIF), a directed fund under the Linux Foundation co-founded by Anthropic, Block and OpenAI, with support from Google, Microsoft, AWS, Cloudflare and Bloomberg. MCP's maintainers continue to run the project day to day under that neutral home. The specification and SDKs live at modelcontextprotocol.io.
How MCP works
MCP messages are JSON-RPC 2.0. A server exposes three main kinds of capability:
create_issue, query_database or search_flights. This is what most people use MCP for.The client lists what the server offers, passes the tool descriptions to the model, and when the model decides to call a tool, the client sends the call to the server and returns the result to the model.
MCP defines two standard transports:
Spec versions
The specification is versioned by date. The main revisions are 2024-11-05 (launch), 2025-03-26 (Streamable HTTP and OAuth authorization), 2025-06-18, 2025-11-25 and 2026-07-28, the first released under AAIF governance and the largest change so far. The 2026-07-28 revision makes the protocol stateless: it retires the initialize handshake and the Mcp-Session-Id header, so each request carries its own protocol version and capabilities, which makes remote servers far easier to run behind load balancers. It also deprecates the Roots, Sampling and Logging features (still working for at least a year) and tightens authorization, including deprecating Dynamic Client Registration in favor of Client ID Metadata Documents. If a client and server disagree on behavior, a version mismatch is the first thing to check.
What people use MCP for
Common problems
MCP and automation
MCP blurs the line between an AI agent and an API integration. An AI agent with the right MCP servers can do much of what a workflow tool does, deciding at run time which tool to call instead of following a fixed graph. In practice the two combine: the agent handles judgment, and MCP servers give it reliable, typed access to the systems it needs.
Autonoly's AI agent can connect to custom MCP servers you add, so tools you or a vendor expose over MCP become available alongside its built-in browser, terminal and native integrations.
Why It Matters
MCP turned 'integrate AI app X with service Y' from an N-by-M problem into building one server per service. Most AI clients and many SaaS vendors now support it, so it is fast becoming the default way to give agents access to real systems, along with the security responsibilities that come with that.
How Autonoly Solves It
Autonoly's AI agent can connect to custom MCP servers, so tools exposed over MCP are available to it alongside its cloud browser, Python terminal and native integrations such as Google Workspace, Slack and Notion.
Learn moreExamples
Adding a GitHub MCP server to an IDE so its coding agent can open issues and read pull requests
A SaaS company shipping a remote MCP server so customers' AI assistants can query their account data with OAuth-scoped access
Connecting an internal MCP server for a company database to an AI agent so it can look up order status while working in a browser
Frequently Asked Questions
Who created MCP and who owns it now?
Anthropic created and open-sourced the Model Context Protocol in November 2024. In December 2025 it donated MCP to the Agentic AI Foundation, a directed fund under the Linux Foundation co-founded by Anthropic, Block and OpenAI. The protocol remains open source, with its maintainers running it under that foundation.
What is the difference between an MCP server and an API?
An API is a service's own interface, with its own URLs, auth and formats. An MCP server wraps capabilities, often an existing API, in a standard shape that any MCP client can discover and call, with names, descriptions and input schemas a model can read. MCP does not replace APIs; it standardizes how AI apps use them.
What transports does MCP support?
Two standard ones: stdio, where the client runs the server as a local subprocess, and Streamable HTTP, where the server is a remote web service. Streamable HTTP replaced the older HTTP+SSE transport in the 2025-03-26 spec revision. Custom transports are allowed but less widely supported.
What changed in the MCP 2026-07-28 specification?
The protocol became stateless: the initialize handshake and the Mcp-Session-Id header were retired, so each request carries its own version and capability information. Roots, Sampling and Logging were deprecated but still work for at least a year, and authorization rules were tightened. Check that your client and server SDKs support the revision before upgrading.
Is MCP safe to use?
The protocol is only as safe as the servers you connect. A server can read and change whatever its credentials allow, and tool descriptions or results can contain prompt-injection text. Install servers from trusted sources, give them least-privilege credentials, and require human approval for destructive actions.
Which apps support MCP?
Most major AI clients: Claude's apps and Claude Code, ChatGPT and OpenAI's agent tooling, IDEs such as VS Code and Cursor, and many agent frameworks and platforms. Support levels differ, especially for remote servers and newer spec revisions, so check the client's documentation.
Further reading
Related terms, automations and guides
Where this concept shows up in practice.
Stop reading about automation.
Start automating.
Describe what you need in plain English. Autonoly's AI agent builds and runs the automation for you — no code required.