Cloudflare Access Forward Proxy (VS Code extension)

A tiny local HTTP forward-proxy that injects CF-Access-Client-Id and CF-Access-Client-Secret headers on every outgoing request. It exists for one specific situation:

You have an OpenAI-compatible endpoint sitting behind Cloudflare Access (service-token gated).
You want a VS Code chat extension to talk to it.
The chat extension lets you change the OpenAI base URL but does not let you add custom HTTP headers.

This extension runs http://127.0.0.1:<port> inside VS Code, transparently forwards every request to https://<your-upstream>:443 with the right CF Access headers attached, and pipes the response back. Auto-starts when VS Code launches.

Includes default-on compatibility shims that make reasoning models (vLLM with Qwen3 / DeepSeek-V4/R1; Ollama 0.3.x+) work cleanly with strict OpenAI clients like Copilot Chat — see vLLM reasoning-model compatibility below.

Build

Requires Node 18+ and npm.

npm install
npm run package

That produces cloudflare-access-proxy-<version>.vsix in the project root.

Install

Local VS Code:

code --install-extension cloudflare-access-proxy-<version>.vsix

Remote-SSH host (install into the remote, not your local UI):

code --install-extension cloudflare-access-proxy-<version>.vsix --remote ssh-remote+<hostname>

Or, with VS Code already attached to the remote, open the Extensions panel and drag-drop the .vsix onto it.

Configure

In Settings UI (search "CF Access Proxy") or settings.json:

{
  "cfAccessProxy.host": "api.nextgarch.org",
  "cfAccessProxy.port": 8787,
  "cfAccessProxy.clientId": "abc123def456.access",
  "cfAccessProxy.enable": true,

  // Reasoning-model compatibility shims (both default true; turn off only if
  // your upstream is a non-reasoning OpenAI-compatible API).
  "cfAccessProxy.disableThinking": true,
  "cfAccessProxy.stripEmptyToolCalls": true
}

Then run CF Access Proxy: Set Client Secret from the command palette (Ctrl+Shift+P) and paste the matching service-token secret. The proxy restarts automatically when settings or the secret change.

Other commands:

CF Access Proxy: Show Status — current host/port/state (also: click the status-bar entry "CF Proxy: :8787").
CF Access Proxy: Restart — stop + start the listener.
CF Access Proxy: Clear Client Secret — remove the stored secret.

Logs live in Output > CF Access Proxy (method/path/upstream status only; no bodies or auth headers).

Point your chat extension at the proxy

Use http://127.0.0.1:8787/v1 (adjust port to match cfAccessProxy.port) as the OpenAI base URL. Any path you hit on the proxy is forwarded verbatim to the upstream, so /v1/chat/completions, /v1/models, etc. all work.

vLLM reasoning-model compatibility

If your upstream is vLLM serving a reasoning model (DeepSeek-V4, Qwen3, R1 family) with --reasoning-parser, two response quirks break strict OpenAI clients like Copilot Chat. Both default-on settings work around them:

cfAccessProxy.disableThinking — when the request is a POST /v1/chat/completions and the caller hasn't set chat_template_kwargs.enable_thinking, the proxy injects chat_template_kwargs.enable_thinking: false into the body. Without this, vLLM emits a long burst of delta.reasoning_content chunks before the first delta.content chunk; Copilot ignores reasoning_content (it's not in the OpenAI spec), concludes the model produced nothing, and shows "Sorry, no response was returned".
cfAccessProxy.stripEmptyToolCalls — when the request carries a tools: [...] array, vLLM emits delta.tool_calls: [] (an empty array) alongside the text content. Strict clients see the presence of tool_calls and treat the delta as a tool-call message, ignoring the content field on the same delta. The proxy removes the empty array from SSE deltas and non-stream message bodies. Real (non-empty) tool calls pass through.

Both are no-ops when the upstream is a stock OpenAI-compatible API, so you can leave them on by default.

For Ollama users: disableThinking also injects top-level think: false, which Ollama 0.3.x+ honors for its reasoning models (Qwen3, DeepSeek-R1, etc.). The two injections coexist harmlessly — each backend ignores the field it doesn't recognize.

Security note

The client secret is stored in VS Code's SecretStorage, which is backed by the OS keyring (Keychain on macOS, libsecret/kwallet on Linux, DPAPI on Windows). It never lands in settings.json and is never written to logs.