What I Learned by Placing GitHub Copilot Behind an MITM Proxy

The author, a developer, was curious about the inner workings of AI-powered applications and chose to investigate GitHub Copilot, as they were quickly exhausting their monthly credits. Noting that many such apps are built on Electron, they used mitmproxy, a popular open-source MITM tool, to intercept HTTP/HTTPS traffic from VS Code. After configuring VS Code to route through the proxy, they observed that Copilot makes several categories of requests even before typing: Auth & Session, Config & Policy, MCP Registry, Repo & Session, Model Discovery, and Recent repos. Before answering prompts, Copilot checks available models via endpoints like /models and /agents/swe/models. With the Auto mode, a request to /models/session/intent classifies the user prompt's intent (code-gen, debugging, reasoning, tool-use) to select the appropriate model. The author tested if secrets from a .env file could be leaked: they added a fake secret to .env and then typed in an unrelated pyproject.toml file. The intercepted inline suggestion request included the secret in the prompt because recently edited files are included as context, demonstrating that disabling Copilot for secret files does not prevent leakage when edits are recent.