Build a custom connector (MCP)

Tool: Python + MCP + VS Code / GitHub Copilot Source status: Advanced connector

An MCP server turns your local spec into callable tools. Instead of pasting context into every prompt, you expose small functions like read_policy(), check_boundary(), or lookup_customer_context().

Steps

Create a Python environment and install MCP.

powershell

python -m venv .venv
.\.venv\Scripts\Activate.ps1
pip install "mcp[cli]"

Create mcp_server.py.

python

from pathlib import Path
from mcp.server.mcpserver import MCPServer

ROOT = Path(__file__).parent
SPEC = ROOT / "spec"

mcp = MCPServer(
    name="local-spec",
    title="Local Spec Tools",
    version="0.1.0",
    instructions="Call these tools before acting on behalf of the user.",
)

def read_spec(name: str) -> str:
    path = SPEC / name
    return path.read_text(encoding="utf-8") if path.exists() else f"[missing: {name}]"

@mcp.tool(description="Return the decision rules. Thin, instant, no model call.")
def decision_rules() -> str:
    return read_spec("rules.md")

@mcp.tool(description="Check whether an action is allowed before it leaves the system.")
def check_boundary(action: str, recipient: str = "") -> str:
    text = f"{action} {recipient}".lower()
    if "external" in text or "customer" in text:
        return "ASK_FIRST\nrule: external communication requires approval\nsource: rules.md"
    if "secret" in text or "personal" in text:
        return "NEVER\nrule: do not disclose sensitive information\nsource: rules.md"
    return "ALLOW\nrule: no boundary matched\nsource: rules.md"

if __name__ == "__main__":
    mcp.run(transport="stdio")

Add a second transport for connector-shaped HTTP calls.

python

import sys

if __name__ == "__main__":
    if "--http" in sys.argv:
        mcp.run(transport="streamable-http", host="127.0.0.1", port=8848)
    else:
        mcp.run(transport="stdio")

Test the local server.
powershell
```
python mcp_server.py
```
Use stdio for VS Code and other local agent hosts. Use streamable-http when you want the same shape a remote connector would expose.

Wire it into VS Code with .vscode/mcp.json.

json

{
  "servers": {
    "local-spec": {
      "command": "python",
      "args": ["mcp_server.py"],
      "cwd": "${workspaceFolder}"
    }
  }
}

Run HTTP mode when you need a connector-shaped endpoint.
powershell
```
python mcp_server.py --http
```

Split tools into thin and thick.

Tool type	What it does	Good for
Thin	Reads files, checks rules, returns structured facts. No LLM call.	Connectors, guardrails, fast grounding.
Thick	Calls another model, runs a council, drafts a long answer.	Local demos and deep reasoning.

Keep connector tools fast

Cowork-style connectors need sub-30-second responses. Put instant file reads and checks behind the connector. Keep slow model calls local or asynchronous.

You'll know it worked when... your agent can call decision_rules() and check_boundary() as tools, and HTTP mode starts on a local port without changing the tool code.

Apply it to your scenario

Expose the smallest stable facts your scenario needs. Good MCP tools are boring and reliable: read the spec, return the current rule, check the boundary, list the known gaps. Let the calling agent do the creative work.

Build a custom connector (MCP) ​

Steps ​

Apply it to your scenario ​

Build a custom connector (MCP)

Steps

Apply it to your scenario