What is MCP?
The Model Context Protocol connects Claude and other clients to your tools, data, and systems through small servers, not one-off custom glue per integration.
MCP (Model Context Protocol) is an open standard for how an AI client (for example Claude Code, Claude Desktop, or other hosts) talks to external capabilities: databases, ticket systems, internal HTTP APIs, Git hosting, files on disk, and more.
Instead of every product inventing its own plugin wire format, MCP defines a shared shape: the host discovers tools (things the model can call), resources (things the model can read or attach), and sometimes prompts (reusable templates). A server implements one domain (for example "our CRM"). The host creates multiple MCP clients, with each client maintaining a one-to-one connection to one server.
Why it matters for coding agents
- Live data: The model can query current issues, schema, or metrics, not only what was pasted into chat.
- Controlled access: You choose which servers run, which tools are allowed, and often which accounts or tokens they use.
- Reuse: The same MCP server can serve multiple clients that speak MCP, not only one vendor's extension format.
In Claude Code, MCP shows up as extra tools in the tool list and as @ resources you can attach (see References with @). Managing connections is typically done through /mcp (or your host's equivalent).
Mental model: host, client, server, capability
| Piece | Role |
|---|---|
| Host | Claude Code (or another app) manages the UX, calls the model, and creates MCP clients. |
| MCP client | One protocol connection inside the host that communicates with one MCP server. |
| MCP server | A small process or service that knows how to talk to one backend (Postgres, Jira, your REST API). |
| Tool | A named action the model can request, with parameters (for example search_issues, run_query). |
| Resource | A readable URI-like thing (often protocol://path) you can @-mention so its data is fetched into model context. Access control and sensitive-data handling remain the host and server’s responsibility. |
The model does not magically get unlimited access: it only gets what the server exposes and what you approved in the host.
Concrete examples
1. Git / code hosting An MCP server might expose tools like "get file from repo", "open pull request", or "list issues". The model can ground answers in real branches and diffs instead of guessing.
2. Database or data warehouse Tools such as "run read-only SQL" (with strict limits) let the model explore schema and sample rows for debugging or migration planning, often safer than pasting full dumps into context.
3. Internal HTTP API
Your company wraps a documented internal API: tools like fetch_customer or create_ticket map to HTTP calls with your auth. The model uses structured arguments instead of inventing URLs.
4. Filesystem or docs root
A server exposes resources under a base path so @ can pull runbooks, OpenAPI specs, or architecture markdown from a known tree.
5. Productivity / ticketing Tools to search Slack-like messages (where supported), create Jira issues, or update Linear, useful for "file a bug from this stack trace" workflows.
Each example is one server (or one product's server package); you enable what you trust.
MCP vs a random script or curl
| One-off script | MCP server | |
|---|---|---|
| Discovery | You paste instructions every session | Host lists tools and schemas for the model |
| Safety | Whatever the script does | Narrow tool surface; host can gate invocations |
| Reuse | Copy-paste between projects | Same server across repos and teammates |
MCP is not a replacement for good permissions in the agent host: combine it with Permissions and Secure execution habits.
See also
Dometrain MCP
Hands-on: add a public HTTP server in Claude Code.
Context7 MCP
Library and framework docs inside the agent.
Playwright MCP
Microsoft Playwright: drive a browser from the agent.
pen.dev MCP
Design on canvas through pen.dev's local pencil server.
MCP in Codex
Codex configuration and client-specific commands.
References with @
Attaching MCP resources in prompts.
Subagents
Agents can use MCP-backed tools when configured.