Skip to content

FAQ.md

Latest commit

 

History

History
131 lines (87 loc) · 5.28 KB

FAQ.md

File metadata and controls

131 lines (87 loc) · 5.28 KB

FAQ

Is this legal? Will my Copilot get banned?

Conduit uses the same Copilot API endpoints that GitHub's own VS Code extension uses. The OAuth Device Flow is the same flow gh auth login uses. That said, GitHub's terms of service for Copilot forbid reselling the capacity — use Conduit for your own personal coding, not as a paid service for others.

At the time of writing (April 2026) there's no public case of an account being banned for local API use, but anyone using this accepts the residual risk.

Does this bypass Anthropic's rate limits?

Conduit doesn't bypass anything — your requests hit Copilot, which has its own per-user limits. If you're a Copilot Business or Enterprise user those limits are generally higher. Heavy thinking usage (Opus 4.7 with big prompts) is what eats the budget fastest.

Why does the Claude Code banner show "Opus 4 · API Usage Billing"?

That banner is a UI cache based on your last Anthropic login, not the live model. The real model being sent is whatever ANTHROPIC_MODEL is set to, which you can confirm in Conduit's dashboard at http://localhost:7023/logs.

Why do I keep getting 401 Invalid API key?

99% of the time: you have ANTHROPIC_API_KEY set in your shell. Claude Code treats that as "Anthropic direct mode" and ignores your ANTHROPIC_BASE_URL. Fix:

unset ANTHROPIC_API_KEY
export ANTHROPIC_AUTH_TOKEN=<your-conduit-key>

Or put unset ANTHROPIC_API_KEY; at the start of your alias.

Can I use this with the Anthropic SDK directly (Python / Node)?

Yes. Set:

from anthropic import Anthropic
client = Anthropic(
    base_url="http://localhost:7133",
    auth_token="<your-conduit-key>",   # NOT api_key
)

Streaming, tool use, thinking, and prompt caching all work.

Can I use this with Cursor / Cline / Aider?

Any tool that lets you set ANTHROPIC_BASE_URL + ANTHROPIC_AUTH_TOKEN (or equivalent) should work. Cline is known to work. Cursor is trickier because it uses its own server-side proxy.

How do I run this in the background?

macOS (launchd)

Create ~/Library/LaunchAgents/com.conduit.proxy.plist:

<?xml version="1.0" encoding="UTF-8"?>
<!DOCTYPE plist PUBLIC "-//Apple//DTD PLIST 1.0//EN" "http://www.apple.com/DTDs/PropertyList-1.0.dtd">
<plist version="1.0">
<dict>
  <key>Label</key>          <string>com.conduit.proxy</string>
  <key>WorkingDirectory</key><string>/path/to/conduit</string>
  <key>ProgramArguments</key><array>
    <string>/opt/homebrew/bin/bun</string>
    <string>run</string>
    <string>dev:proxy</string>
  </array>
  <key>EnvironmentVariables</key><dict>
    <key>CONDUIT_API_KEY</key><string>your-key-here</string>
  </dict>
  <key>RunAtLoad</key><true/>
  <key>KeepAlive</key><true/>
  <key>StandardOutPath</key><string>/tmp/conduit.log</string>
  <key>StandardErrorPath</key><string>/tmp/conduit.err</string>
</dict>
</plist>

Then launchctl load ~/Library/LaunchAgents/com.conduit.proxy.plist.

Linux (systemd)

Create ~/.config/systemd/user/conduit.service:

[Unit]
Description=Conduit proxy
After=network.target

[Service]
WorkingDirectory=/path/to/conduit
ExecStart=/home/you/.bun/bin/bun run dev:proxy
Environment="CONDUIT_API_KEY=your-key-here"
Restart=always

[Install]
WantedBy=default.target

Then systemctl --user enable --now conduit.

Does prompt caching work?

Yes. Conduit preserves per-block cache_control in the passthrough, and Copilot's backend honors it. The cache_creation_input_tokens and cache_read_input_tokens fields come back correctly in the usage object. You can see cache hit rates in the dashboard.

I have a model that isn't on Copilot. Can Conduit route to it?

Yes — the "Custom upstream providers" feature. Open the dashboard, go to Settings → Upstream Providers, and add an OpenAI-compatible endpoint. Any request whose model matches a configured prefix will be routed to that provider instead of Copilot.

The -1m thing — do I need anthropic-beta: context-1m-*?

For claude-opus-4.6 specifically, no. We've measured it accepting up to 1,000,000 input tokens without any beta header. The -1m variant exists in Copilot's catalog but the base variant is already 1M-capable. Conduit will still switch to -1m if you pass the beta header, for future-proofing.

For other models the limit is lower (168K for most Sonnet/Opus 4.5+, 136K for Haiku 4.5).

What about Anthropic's 200K / 400K / 1M pricing tiers?

Copilot flat-rates the underlying inference as part of your subscription, so tier pricing doesn't apply here. You're paying GitHub, not Anthropic.

Streams get cut off on long thinking responses. What's going on?

Bun's default HTTP idle timeout is 10 seconds. Conduit raises it to 255 (Bun's max). If you still see cut-offs, it's almost always the upstream (Copilot) silently killing slow streams — in which case retrying with a shorter effort or splitting the prompt usually fixes it.

SSE keepalive (periodic : keepalive\n\n comments) is on the roadmap to make this fully bulletproof.

How do I contribute?

PRs welcome. Things that would actually help:

  • Tests for edge cases we've missed (report them as issues first)
  • SSE keepalive implementation
  • DB-backed multi-tenant API keys
  • Better Cursor / Cline / Aider compatibility notes

Run bun run typecheck && bun run test before opening a PR.