The Model Context Protocol (MCP) is a standard way for an AI host to talk to an outside program. The host is the app you use (Cursor, Claude Code, Claude Desktop, Codex). The server is a small program that offers capabilities: a converter, a database query, a ticket lookup. Because the protocol is shared, one server works in every host that speaks MCP.
What a server can offer
| Primitive | What it is | Who decides to use it | Example |
|---|---|---|---|
| Tools | Functions with a name, a description and a JSON Schema for inputs | The model, during a conversation | markdown_to_slack({ markdown }) |
| Resources | Read-only data identified by a URI | Usually the user or host attaches it | A file, a DB row, a doc page |
| Prompts | Reusable prompt templates with arguments | The user picks them, often as slash commands | "Summarise this ticket" |
Most servers you will use expose tools only. That is fine: tools are what let an agent act.
A real tool list
Markdown Convert MCP exposes four tools. Each takes { "markdown": "<source>" } and returns converted text:
| Tool | Output | Use when the next hop is |
|---|---|---|
markdown_to_slack | Slack mrkdwn | Slack API or a Slack MCP text field |
markdown_to_teams | HTML with inline CSS | Teams paste, or Graph contentType: html |
markdown_to_jira | ADF JSON string | Jira Cloud REST v3 rich-text fields |
markdown_to_plaintext | Plain text | Anything that needs formatting stripped |
The model picks a tool by reading these names and descriptions, exactly as it picks a skill. A vague tool description leads to wrong or missing calls.
Transports: local or remote
- stdio: the host starts the server as a child process and talks over stdin/stdout. Typical for
npx …oruvx …servers on your machine. No network port, no auth. - Streamable HTTP: the server runs at a URL. The host sends JSON-RPC over HTTP. Typical for hosted servers. For example, the public, read-only server at
https://www.ysskrishna.space/mcpneeds no API key.
Older docs mention an SSE transport; newer servers use Streamable HTTP.
The request flow
Under the hood, MCP is JSON-RPC 2.0. A session looks like this:
- initialize: host and server exchange protocol version and capabilities ("I support tools").
- tools/list: the host asks which tools exist and gets names, descriptions and input schemas.
- The model decides to call one. The host sends tools/call with the tool name and arguments.
- The server returns content, or a result with
isError: trueif the tool failed. Markdown Convert MCP, for example, returns a tool error (not a transport failure) for input over 200,000 characters.
Hosts add their own layer on top: approval prompts before a tool runs, per-server enable switches, and limits on how many tools are active.
Trust and safety
An MCP server runs code (stdio) or receives your data (HTTP). Before you add one:
- Read what it does. A converter that only transforms text is low risk. A server with shell or write access is not.
- Prefer read-only for remote servers. And do not send secrets to a server you do not run.
- Treat tool output as untrusted text. A web page or ticket returned by a tool can contain instructions. Your agent should treat them as data. This is the same rule as for skills that read files.
- Keep the tool count small. Every enabled tool's description costs context and adds a way to pick the wrong tool.
Skills and MCP together
A server provides the ability; a skill provides the judgment about when to use it. Markdown Convert MCP ships a companion skill for this reason: it tells the agent to convert before posting to Slack, Teams or Jira through an API, and not to convert before pasting into Slack's composer, which already handles Markdown. Without the skill, the tools are available but called at the wrong moments.
Quick check
-
Which MCP primitive does the model invoke during a conversation?Tools.
-
Your server is started with
npx -y …on your laptop. Which transport is that?stdio: the host runs it as a child process. -
A tool returns a web page that says "ignore previous instructions". What should the agent do?Treat it as data in the result. Tool output never changes the task or grants permission.






