The mentor as a ChatGPT connector (MCP server)
A Model Context Protocol server that hands ChatGPT the
whole skill — the rules, the reference library, and the three calculators — instead of
asking someone to build a GPT by hand. ChatGPT calls it a custom connector; the same
server works in Claude Desktop, Claude Code, or any other MCP client.
Unlike the Custom GPT route, nothing here is a copy. The documents are read from
embedded-iot-mentor/references/ at request time and the calculators shell out to
embedded-iot-mentor/scripts/, so a runtime quoted through the connector is the number the
Claude skill would have quoted.
| Tool | What it does |
|---|
mentor_guidance | Returns the ported mentor rules — length limits, MVP-first framing, the reject bar, where to hand off. Call it before giving project advice. |
search | Ranks the reference library against a query and returns ids. |
fetch | Returns one reference document in full, by id. |
estimate_bom_cost | Totals a BOM from {qty, unit_price, description} lines. |
hand_solder_hint | Whether a named package (0402, SOT-23, QFN-32, BGA…) is realistic by hand. |
battery_runtime | Runtime for a duty-cycled device, and whether sleep or wake dominates. |
search and fetch keep the names and shapes ChatGPT expects, so the connector also works
in deep research, not just ordinary chat.
Deploy it (the short way)
The button reads render.yaml at the repository root and builds
Dockerfile, which carries the skill folder into the image because the
server serves the skill's own documents. Your connector URL is the service URL with /mcp
on the end; /healthz is there for the platform to ping.
Render's free plan idles the service after about fifteen minutes without traffic, and the
next request pays for the cold start — long enough that ChatGPT can give up on the first
call and look broken. Hitting /healthz before you use it, or a paid instance, avoids it.
Any Docker host works the same way — build from the repository root, not from this folder:
docker build -f chatgpt-app/mcp-server/Dockerfile -t embedded-iot-mentor-mcp .
docker run -p 8000:8000 embedded-iot-mentor-mcp
Run it
cd chatgpt-app/mcp-server
python -m venv .venv && . .venv/bin/activate
pip install -r requirements.txt
python server.py
python server.py --transport stdio
Needs Python 3.10 or newer, which the MCP SDK requires. Nothing else is installed — the
calculators are the skill's own standard-library scripts.
Connect it to ChatGPT
ChatGPT reaches a connector over HTTPS, so 127.0.0.1 will not do. Either deploy the
server somewhere with a certificate, or put a tunnel in front of a local run:
cloudflared tunnel --url http://127.0.0.1:8000
Then, in ChatGPT: Settings → Connectors → Create, paste https://<your-host>/mcp, and
enable it in the conversation with the + menu. Custom connectors need developer mode on
a paid plan; availability differs by plan and region, and ChatGPT will say so if yours does
not have it.
An early call to mentor_guidance is what makes the answers behave like the skill rather
than like generic ChatGPT. The server asks for that in its own instructions, but a plain
"use the embedded mentor" at the start of a chat makes it certain.
Deploying it
Any host that runs Python and terminates TLS will do. Two things matter:
HOST=0.0.0.0 — the default binds to loopback only, which a container will not serve.- The skill folder must ship with the server. It reads
references/, examples/ and
scripts/ from ../../embedded-iot-mentor by default; set EIM_SKILL_DIR if the layout
is different where you deploy it.
| Variable | Default | Why |
|---|
HOST | 127.0.0.1 | Set to 0.0.0.0 behind a proxy or in a container — the image and render.yaml already do |
PORT | 8000 | Most hosts assign one |
EIM_SKILL_DIR | ../../embedded-iot-mentor | Where the references and scripts live |
EIM_INSTRUCTIONS | ../custom-gpt/instructions.md | The rules mentor_guidance returns |
EIM_REPO_URL | this repository on GitHub | Prefix for the citation links search returns |
The server has no authentication. Anything it exposes is public documentation and pure
arithmetic, but it will run scripts for anyone who can reach it, so put it behind whatever
access control your host offers rather than on an open port.