LiteLLM Proxy for GitHub Copilot

A Nuanst extension. Route GitHub Copilot — chat, code completions, inline hints/NES, agents, and skills — through your LiteLLM proxy or local models, and keep Copilot as the UI. Useful when model traffic should stay on your own infrastructure, including after Copilot credits are exhausted.

Install

Search the Extensions view for LiteLLM Proxy for GitHub Copilot (publisher Nuanst-GmbH), or install a packaged build:

code --install-extension vscode-litellm-0.2.1.vsix

Reload the window after installing. You need VS Code 1.110 or newer, GitHub Copilot Chat, and a LiteLLM proxy (default http://localhost:4000).

How it works

github.copilot-chat funnels all model traffic through two base URLs and ships hidden settings that override them (verified against the 0.38 bundle):

This extension writes all six key forms for you (current + legacy prefixes, see OVERRIDES in src/copilotSettings.ts), manages the Copilot-model → LiteLLM-deployment mapping, health-checks the proxy, and can export a ready litellm/config.yml. Auth token minting still goes to GitHub by design — see docs/spike.md for the credits-exhaustion findings.

Setup

1. Set vscode-litellm.proxyBaseUrl (default http://localhost:4000) and start your LiteLLM proxy.
2. If the proxy requires an API key, run LiteLLM Proxy: Store API Key Securely and turn on vscode-litellm.forwarder.enabled. Copilot sends its own token; the forwarder replaces it with yours.
3. The status bar shows ✓ LiteLLM when the proxy is reachable, and ✗ when it is not. Click it for every command.
4. Commands:
   • LiteLLM Proxy: Configure Model Map — map Copilot model names → LiteLLM deployments
   • LiteLLM Proxy: Test Connection
   • LiteLLM Proxy: Recent Requests — last 200 forwarded requests with status, latency, time-to-first-byte, retries and errors (no more log-scrolling to find the failed call)
   • LiteLLM Proxy: Store API Key Securely / Clear Stored API Key — keep the proxy key in VS Code SecretStorage instead of settings.json (SecretStorage takes priority while set)
   • LiteLLM Proxy: Export LiteLLM config.yml — generates a config using the Copilot model names as model_name
   • LiteLLM Proxy: Toggle Routing — on: override keys written; off: those keys cleared and the real GitHub endpoints restored

Settings

Limitations

Turn on the forwarder when the proxy checks an API key. With it on, the extension also covers two Copilot paths LiteLLM does not:

• POST /v1/engines/<model>/completions (inline completions) is rewritten to /v1/completions.
• GET /models is answered in Copilot's shape. GET /v1/models stays OpenAI-shaped, with your mapped names merged in.

GET /agents is answered with an empty list. Other Copilot-only routes (/skills, /search, /models/session) are forwarded as-is and depend on what your proxy implements.

Troubleshooting

• Copilot still hits GitHub after toggling: reload the window; the overrides re-apply on config change and on every Copilot token refetch. Check that the override keys exist in global settings.json (command palette → Preferences: Open User Settings (JSON)).
• After a Copilot update routing stops: re-run Toggle Routing off→on (the hidden keys are undocumented and could change name or behavior in a future Copilot build — the extension writes both current and legacy forms).
• ✗ in status bar: run Test Connection; it reports which endpoints were tried. LiteLLM must be reachable over plain HTTP/HTTPS from the extension host.
• Rate-limit/429 noise from Copilot: the proxy may omit x-ratelimit-* / x-github-request-id headers; Copilot tolerates their absence.
• "which Copilot call just failed?": LiteLLM Proxy: Recent Requests lists the last 200 forwarded requests with status, total latency, time to first byte, and whether a fallback/transport retry happened; pick one for the full detail. The forwarder also transparently retries once on connection errors and on 429/5xx (before any byte reaches the client), so a flaky proxy blip may not produce a visible error at all.
• Copilot updated and routing silently stopped: the doctor now compares the running VS Code / Copilot Chat build against the verified-build table (src/copilotVersions.ts). If it warns the build is unverified, re-run the spike and add the build to the table.
• Language model unavailable (log: Could not find model 'X' in local cache after re-resolving models): the selected model isn't in the proxy's model list, so Copilot rejects it before any request is sent. Fix: enable forwarder.enabled, set forwarder.fallbackModel to a model the proxy actually serves, and add the selected name (e.g. copilot/auto) to the model map — the forwarder then advertises that name in the model list and retries refused requests on the fallback model.
• "I don't see any chat traffic in the log": at the default trace level the forwarder logs every real client request as POST /chat/completions (…B) → <upstream> + ← <status>. The recurring GET /v1/models / GET /models lines are mostly the extension's own 30-second status poll (tagged x-litellm-router: status-poll, logged at debug) and Copilot's model-list fetch — they are not a chat turn. A chat turn only exists once you actually submit a prompt. Run Troubleshoot (doctor) → it sends one 1-token POST /chat/completions through the forwarder and reports the upstream status, and it probes that a WebSocket upgrade is refused cleanly (501) rather than hanging — a silent WS hang looks exactly like "chat isn't routed".
• "message filter" line in the log (dropped N empty system message(s), nulled M empty assistant content(s)): the forwarder normalized Copilot's blank system prompt / empty tool-call assistant content before forwarding. Without it, LiteLLM's modify_params rewrites those empty strings into a visible "[System: Empty message content sanitised…]" placeholder that persists in the conversation history.

Development

npm install
npm test
npm run compile          # type-check, then esbuild → dist/extension.js

Press F5 in this workspace to launch an Extension Development Host. Packaging and Marketplace steps are in PUBLISHING.md. The endpoint checklist for a new Copilot build is docs/spike.md.

Repo layout

src/extension.ts        activation, commands, config validation + wiring
src/copilotSettings.ts  writes/clears the hidden override keys (API + settings.json fallback)
src/copilotVersions.ts  verified VS Code / Copilot Chat builds for the override keys (doctor check)
src/forwarder.ts        local auth-rewriting proxy: advertises modelMap names, fallback-model retry, transport retry (429/5xx/connect errors), first-byte timeout, empty-content normalization, recent-request ring buffer
src/policies.ts         pure forwarder decision logic (unit-tested, no vscode)
src/configRender.ts     LiteLLM config.yml rendering (pure, unit-tested)
src/proxy.ts            LiteLLM health/model checks
src/modelMap.ts         quickpick model-map editor
src/statusBar.ts        status bar + configurable polling
src/notify.ts           deduped warning toasts
test/                   standalone smoke tests (npm test, no framework)
spike/config.example.yml  minimal LiteLLM config for the spike
docs/spike.md           checklist for verifying override keys after a Copilot update

License

MIT © 2026 Nuanst One GmbH.

GitHub Copilot is a trademark of GitHub, Inc. LiteLLM is a trademark of BerriAI. This extension is not affiliated with or endorsed by either organization.