MCP Connectors and Integrations
What a connector is
A Model Context Protocol (MCP) connector is the structured interface that links an AI agent to an outside system. A connector — implemented as an MCP server — exposes a catalog of tools, resources, and prompts, each with a declared schema. That schema is what lets an agent discover a capability and call it safely, without embedding API logic or credentials in the model itself.
Connectors exist for local utilities (files, a local database) and for remote platforms (hosted databases, source control, CMS APIs). Collectively they form the fabric that makes agentic AI practical: the model reasons, the connector acts.
For the roles and transports referenced below, see MCP Foundations and Architecture.
Anatomy of a connector
| Layer | Purpose | Examples |
|---|---|---|
| Server | Implements the MCP endpoints and defines the tools and resources. | supabase-mcp, chrome-devtools-mcp |
| Descriptor | A JSON entry that tells the client how to launch or reach the server. | mcp.json, a client’s mcpServers config |
| Transport | How messages flow — stdio for local, HTTP for remote. |
npx @supabase/mcp or https://mcp.supabase.com/mcp |
| Host integration | How the user’s AI environment registers and uses the connector. | Claude Desktop config, Cursor settings |
All traffic is JSON-RPC 2.0 (tools/call, resources/read) against declared JSON schemas, which is what gives connectors cross-client portability.
Two integration patterns
Local (stdio). The client launches the server as a subprocess via a CLI command (npx, uvx) and talks to it over standard I/O. Best when an agent inside a local IDE needs the filesystem or a local tool. Strengths: privacy and direct system access, no network exposure.
Remote (HTTP). The server runs behind a public https://.../mcp endpoint. Best for cloud services and shared enterprise APIs. Strengths: broad client compatibility and centralized operation. Requires real authentication (OAuth 2.1 with scopes).
Example connectors
| Connector | Purpose | Integration notes |
|---|---|---|
| Supabase MCP | Database and auth access for agents. | Remote HTTP; OAuth 2.1; grouped by feature. |
| Chrome DevTools MCP | Browser debugging and performance analysis. | Uses the DevTools protocol; suited to developer agents. |
| WordPress.com connector | Read access to WordPress.com site data. | Official integration over OAuth 2.1. |
| Context7 Memory MCP | Persistent agent memory and retrieval context. | Multi-agent context sharing; vector queries. |
| Google Developer Knowledge MCP | Access to official Google developer docs. | Covers Firebase, Android, Google Cloud; needs a Google Cloud key. |
SDKs for building connectors
You rarely hand-write the JSON-RPC loop. Established SDKs handle transport, schema validation, and lifecycle:
| SDK | Language | Notes |
|---|---|---|
| FastMCP | Python | Lightweight framework for quickly standing up a server. |
| MCP PHP SDK | PHP | Framework-agnostic; suits WordPress and Symfony/Laravel stacks. |
| MCP TypeScript SDK | TypeScript | Builds npm-published servers such as the Chrome DevTools MCP. |
Each abstracts the boilerplate and provides helpers for declaring tools, validating schemas, and managing the server’s startup and shutdown.
Discovery and security
- Discovery is dynamic: a client calls
tools/listandresources/listto learn a server’s capabilities at connect time, rather than being compiled against them. - Authentication differs by transport. Remote servers should require OAuth 2.1 with scoped tokens; local servers inherit the system user’s trust boundary.
- Auditing belongs to the Host: record the full JSON-RPC exchange so tool calls can be reviewed and governed.
For the full security stack — token structure, sandboxing, and compliance — see MCP Security and Compliance.
Designing a good connector
- Keep scope tight. One server, one clear domain (analytics, CMS, storage). Narrow surfaces are easier to reason about and secure.
- Version explicitly. Use semantic versioning on the server and its tools so clients can handle breaking changes.
- Write schemas an agent can read. Explicit types and plain-language descriptions measurably improve tool-selection accuracy.
- Fail with structure. Return typed errors with codes, not free text, so an agent can self-correct.
- Default to read-only. Mark retrieval tools
readonly; reserve state-changing tools for deliberate opt-in. - Gate mutations. Require a human confirmation step for any operation that writes or deletes.
