Why it mattersA pool of personal subscriptions can now behave like one provisioned provider, so a small team can raise its daily agent quota without moving off subscription pricing to metered API billing.
Geoffrey Huntley on Saturday published underclass, a local proxy that presents one OpenAI-compatible endpoint and fans requests across a pool of personal ChatGPT Codex and GitHub Copilot subscriptions. The Rust project is MIT-licensed and was created on 20 September at 07:00 UTC. It had 97 stars, 5 forks and one open issue when this piece was written, roughly twelve hours after the first commit.
Huntley describes it in the README as turning "a pile of personal subscriptions into a single durable endpoint that behaves like one well-provisioned provider". The proxy speaks /v1/responses, /v1/chat/completions and /v1/models, so any OpenAI-compatible client can be pointed at it with no code changes. Accounts are added through the web UI using the vendor's own OAuth device flow, one at a time.
Sticky sessions keep the prompt cache warm
The routing rule that matters for a coding agent is stickiness. When a request carries a prompt_cache_key, the proxy always sends it to the same account, so the upstream prompt cache keeps working across turns. The README says opencode emits this key when configured with setCacheKey: true. Bindings live for 24 hours and survive restarts.
A quota response moves the account it came from into a cooling state until its Retry-After window passes, and it returns to rotation automatically. If every eligible account is cooling, the proxy answers 429 with the earliest reset time. It never queues. Codex and Copilot accounts compete flat by least-in-flight, filtered by each backend's model catalog.
The pool includes GitHub Copilot subscriptions
The Copilot support is the part that broadens the audience. GitHub Copilot Pro subscribers can feed their Copilot quota into the same pool as their Codex quota, through the same endpoint, and see the union of both catalogs at /v1/models. The model catalog is a configuration file, seeded with the current Codex families named in the README plus Copilot's live list, and editable from the UI or the admin API.
Access tokens, authorization headers and prompt bodies are never logged, according to the security section. Account labels (an email or username) appear in logs and the UI by design; raw account IDs are truncated. auth.json is written with mode 0600 and the proxy binds to localhost by default.
Underclass also correlates with Huntley's earlier preflight proxy through an x-request-id header, so a request can be traced through inspection and pool routing with the same ID.
The practical change is small and specific: a group of independent contractors, each holding a personal Codex or Copilot subscription, can serve all their quotas through one endpoint and let a shared agent send each request to whichever account is currently available. Two clients running against the same pool never see a per-account rate limit until every account in the pool is cooling. Each subscription's terms of service still apply to each account individually, and each vendor's usage policy is unchanged by fronting the accounts behind a proxy.
Source
Source: GitHub
This item was written by an AI system from the linked source. Reveneau is responsible for what it publishes.
Get AI News in your inbox
New developer tools, model and agent releases, and how teams are actually using them to release software. Short, and only when there is something worth reading.