use-someone-elses-subscription

Run claude -p, codex exec, or agy -p on your machine, over HTTPS, for someone else.

Starts a tiny local HTTP server in front of the CLI and punches a public URL to it through a Cloudflare quick tunnel. Anyone with the URL and a bearer token can send it a prompt and get back whatever the CLI printed — running on your machine, under your login/subscription.

No API key sharing, no cloud account, no server to rent. Just your laptop/VM and a tunnel.

Requirements

• Whichever of claude, codex, or agy you want to serve — installed and logged in
• cloudflared on your PATH
• uv

Install

uv tool install git+https://github.com/MT-GoCode/use-someone-elses-subscription

Use

give-sub claude start   # runs claude -p, default port 8787
give-sub codex start    # runs codex exec, default port 8788
give-sub agy start      # runs agy -p, default port 8789

All three can run at once (separate ports/tunnels/state).

give-sub (claude) is live:
  URL:   https://some-random-words.trycloudflare.com
  Token: 3f9a...

curl -X POST https://some-random-words.trycloudflare.com/v1/chat/completions \
  -H 'Authorization: Bearer 3f9a...' \
  -d '{"messages":[{"role":"user","content":"hi"}]}'

Base URL for the HTML client / OpenAI-compatible clients (token baked in): https://some-random-words.trycloudflare.com/t/3f9a.../v1

Send the URL + token to whoever you're giving access to. They can call it with plain curl, or open use-someone-elses-subscription.html in a browser and paste in that last "Base URL" line — nothing else to configure.

give-sub claude status   # show the current URL/token, if running
give-sub claude stop      # kill the server + tunnel

(give-sub is the short alias; use-someone-elses-subscription also works, same subcommands.)

API

POST /v1/chat/completions — same shape as OpenAI's chat completions endpoint, so existing OpenAI-client code can point base_url at your tunnel URL.

• Auth: either Authorization: Bearer <token> header, or POST /t/<token>/v1/chat/completions with no header — for clients that only offer a single "Base URL" field, use https://host/t/<token>/v1 as that URL.
• Body: {"messages": [{"role": "user", "content": "..."}]} — the last message's content becomes the prompt
• Response: a minimal OpenAI-shaped completion, with choices[0].message.content set to whatever the backend CLI printed

There's no multi-turn memory, no streaming, and no filesystem access — every call runs in a fresh empty temp directory and is capped at 2 concurrent requests (see GIVE_CLAUDE_MAX_CONCURRENT / GIVE_CLAUDE_TIMEOUT env vars).

This does not work as the base URL for the real Claude Code / Claude Desktop app, the real Codex app, or the real agy app — those speak their provider's actual streaming API, not this stripped-down OpenAI shape.

Careful with this

• Whoever has the token can spend your usage/quota on that backend. Don't post the token+URL somewhere public.
• The quick tunnel URL is random and dies with the process — restart it and the URL changes. Fine for handing to one person for a while; not meant to be a stable, permanent service.