How it works
MCP defines three roles: hosts, clients and servers. The host is the AI app. The client is the connector inside it. The server is the tool: it lists what it can do and does it when asked. Messages are JSON-RPC 2.0 messages.
client → server tools/list server → client get_weather: "Get current weather information for a location" client → server tools/call get_weather {"location": "New York"} server → client "Current weather in New York: 72°F, partly cloudy"
The model reads the tool list, picks a tool and fills in its arguments. The app sends the call. The spec says there should always be a human in the loop who can deny a tool call, so apps usually ask first.
What a server offers
A server can offer tools, resources and prompts.
| Tools | Resources | Prompts | |
|---|---|---|---|
| What it is | Functions the model can run | Context and data to read | Ready-made message templates |
| Who picks it | The model, from the descriptions | You or the model | You |
| Example | Run a command, create a computer | A file, a database row | "Review this change" |
Local or remote
The spec defines two standard transports: stdio and Streamable HTTP. With stdio, the app starts the server as a program on your own machine. With Streamable HTTP, the server runs somewhere else and each message is an HTTP POST. To connect a remote server in Claude Code:
# replace the name and the address with your server's $ claude mcp add --transport http example https://example.com/mcp $ claude mcp list
The spec is versioned by date, and its current revision is 2026-07-28. In that revision the current revision has no protocol-level session; every request carries its own protocol version and capabilities.
MCP or a plain API?
| A plain API | MCP | |
|---|---|---|
| Made for | Programs a person wrote | AI models and the apps around them |
| How you learn what exists | You read the docs | The server lists its tools, each with a name, a description and the arguments it takes |
| One integration works in | Your own code | Any app that speaks MCP |
They work together: many MCP servers wrap an existing API, so the model gets named tools instead of raw web requests.
Do you need it?
Use MCP when
- An agent has to act in other systems. Files, a calendar, a database, a hosting account.
- You want one tool in many AI apps. Build one server instead of one plug per app.
- You want the model to find tools by itself. It reads the list and chooses.
Skip it when
- You run one fixed script. Call the API from your own code.
- The tool has no MCP server yet. A short script the agent can run does the same job.
- You would not trust the server with your data. Connect only servers you trust.
What to watch
- The spec treats tools as arbitrary code execution. Read what a tool does before you allow it, and treat a server's own descriptions with caution.
- An agent with many tools can pick the wrong one or call one too often. Connect the few it needs.
- MCP is young and changes: the revision on this page is 2026-07-28. A server built for an older revision may need an update.
MCP on Uno
Uno is driven by API and MCP tools. An agent can create a computer by MCP and work on it, and the computer sleeps when the job is done. Your own AI can connect to Uno over MCP, so the agent that plans the job is also the one that sets up where it runs.
Your agent gets a clean address of its own
Your computer runs in the Netherlands. Its traffic can leave through a static address of a regular internet provider in one of 8 countries. Only your account works from that address.
- From a data-center address
- Sites see a server range shared with thousands of strangers. The agent's browser hits more captchas and blocks.
- From your Clean exit IP
- Sites see a regular internet provider's address in the country you picked. Fewer captchas, and less reason for Claude or ChatGPT to flag your account: nobody else works from it.
The short way: an agent that sets up its own computer
Point your agent at Uno and ask for a computer. It gets a Linux computer with root and Docker, and the computer sleeps when the job is done.
Small: $5 a month
A computer your agent can create, from $5 a month
Every plan is a Linux microVM with root and Docker. Move up or down any time; your files come with you.
Small
A Telegram bot, a VPN, a small backend
$5/ month
- 1 vCPU · 1 GB
- 15 GB disk
- 25 GB cloud storage
- 5 boosts a month
Plus
Agents that work with your laptop closed
$20/ month
- 4 vCPU · 4 GB
- 40 GB disk
- 100 GB cloud storage
- Uno Work in the cloud
Pro
A product in production, a second MacBook
$70/ month
- 16 vCPU · 16 GB
- 120 GB disk
- 300 GB cloud storage
- Speed of a MacBook Air M4
Max
A team and many agents at once
$200/ month
- 32 vCPU · 48 GB
- 300 GB disk
- 1,024 GB cloud storage
- 5 public IPv4
Add Uno AI to any plan: 40 h of AI work a month on Small, 100 h on Plus, unlimited on Max. Today: GLM-5.3 Flash and MiMo v2.6 Pro. Or sign in with your own Claude or ChatGPT.