All capabilities Agents, tools and MCP

Give a model your own tools and data through MCP

One open protocol between an AI application and your systems, so a connector you write once works in every client that speaks it.

DocsOfficially documentedTestNot tested by benchr Medium · API Docs checked September 1, 2026

Claude CodeClaude APIMCP serversOpenAI API+3

Overview

Standardises how an application exposes data sources, tools and workflow prompts to a model.
Is supported across clients rather than one vendor - the spec site names Claude, ChatGPT, VS Code, Cursor and others.
Lets a coding agent reach a design file, a ticket tracker or an internal database without a bespoke integration per client.

01Decide what the model actually needs: a read-only resource, a callable tool, or a prompt template. Most first servers should be read-only.
02Build the server against the current spec revision and test it in one client before wiring it into an agent loop.
03Scope credentials to the server, not to the model. The server is the trust boundary.
04Treat every value that comes back through a server as untrusted content, the same way you would treat a web page.

MCP standardises the connection, not the safety of what you connect. A server with broad write access hands that access to whatever is driving it.
Client support varies by feature; a server that works in one client is not automatically fully featured in another.
The spec is versioned and revises - the docs are organised by dated revision, so pin the one you build against.

benchr has not run this capability. No test has been published.

benchr test runs →

Source-check history

September 1, 2026DocsOfficially documentedFirst logged from the protocol's own introduction, including the named client list and the dated spec revision structure.