AI Gatekeeper

A local security proxy that sits between your AI coding assistant and the internet, intercepting requests in real time to detect and block sensitive data leaks before they leave your machine.

Built with mitmproxy and a local LLM (Ollama), it requires no cloud services and adds zero latency to allowed requests.

(docs/architecture.png)

---

How it works

Every request VS Code sends to GitHub Copilot is intercepted by the proxy. The payload is parsed to extract the actual user query, then checked in two phases:

1. Regex scan — instantly flags API keys, tokens, and credentials matching known patterns (OpenAI keys, GitHub tokens, AWS access keys, generic secrets)
2. LLM policy check — a local Llama model evaluates the request against a custom policy defined in instructions.md

Requests that pass both checks are forwarded to Copilot. Anything that fails returns a 403 Policy Violation with the reason logged.

---

Requirements

• macOS (tested) / Linux
• Python 3.9+
• mitmproxy
• Ollama with llama3.2:1b pulled

---

Setup

1. Install dependencies

brew install --cask mitmproxy
pip install mitmproxy requests
ollama pull llama3.2:1b

2. Trust the mitmproxy certificate

Start mitmproxy once to generate the certificate, then trust it:

mitmproxy  # run briefly, then quit with q
sudo security add-trusted-cert -d -p ssl -p basic \
  -k /Library/Keychains/System.keychain \
  ~/.mitmproxy/mitmproxy-ca-cert.pem

3. Configure VS Code

Add to your settings.json:

{
  "http.proxy": "http://127.0.0.1:8080",
  "http.proxySupport": "override",
  "http.proxyStrictSSL": false,
  "http.systemCertificates": true
}

4. Define your policy

Edit instructions.md to set what the LLM should block. Example: Do not allow sharing of internal project names, credentials, or any personally identifiable information.

---

Running

Open two terminals:

# Terminal 1 — start the local LLM
ollama serve

# Terminal 2 — start the proxy
python3 start_gatekeeper.py

The proxy listens on http://127.0.0.1:8080. All Copilot requests from VS Code will now be intercepted and checked automatically.

---

Test mode

Set TEST_MODE=true (default) to intercept all matched requests and return a mock response instead of forwarding to Copilot. Useful for testing your policy rules without consuming API quota.

TEST_MODE=true python3 start_gatekeeper.py   # no real requests sent
TEST_MODE=false python3 start_gatekeeper.py  # live mode

---

Project structure

ai-gatekeeper/ ├── gatekeeper.py # core proxy logic — request parsing, regex, LLM check ├── start_gatekeeper.py # launcher — starts mitmdump with gatekeeper.py ├── instructions.md # your policy file — edit this to customise rules ├── requirements.txt └── docs/ └── architecture.png

---

Verifying interception (optional)

To confirm the proxy is intercepting traffic before enabling the security layer:

./launch-with-mitmproxy.sh code .

This opens VS Code with mitmproxy running in a separate terminal so you can inspect raw traffic.

To manually test a proxied request:

curl --proxy http://127.0.0.1:8080 "http://wttr.in/?0"