aisdk-pi-proxy

A tiny Hono HTTP server that accepts OpenAI Responses API requests (e.g. from @ai-sdk/openai's openai.responses(modelId)) and forwards them to any provider supported by @mariozechner/pi-ai. Defaults to OpenAI Codex (ChatGPT Plus/Pro OAuth).

The Responses API has native event types for reasoning summaries, structured output items, and function-call argument streaming — pi-ai's unified events map to it directly with no <think> tag smuggling and no AI-SDK middleware.

• Reasoning is native. pi-ai's thinking_* events become response.reasoning_summary_text.delta chunks; AI SDK surfaces them as proper reasoning-* parts.
• Auth is provider-agnostic. Swapping PI_PROVIDER switches between Anthropic Pro/Max OAuth, GitHub Copilot, Gemini CLI, Antigravity, OpenAI Codex, or any API-key provider — pi-ai handles refresh.
• Reasoning effort per request. Send x-pi-reasoning: minimal|low|medium|high|xhigh to override the default; the request body's reasoning.effort is also honored.

⚠️ ToS warning

Using a ChatGPT Plus/Pro subscription (or Anthropic Pro/Max, etc.) for programmatic traffic likely violates the provider's terms. Realistic failure modes: rate-limiting, account suspension, silent capability downgrades, endpoint shape changing without notice. Personal/dev use only — do not ship this to other users.

Prereqs

1. Install pi-ai's CLI and log in once on the host:

   cd aisdk-pi-proxy
   npm install
   npx @mariozechner/pi-ai login openai-codex

   This writes auth.json in the current directory. The proxy reads from the same path on startup; refreshes are written back.

2. Other providers (any of anthropic, github-copilot, google-gemini-cli, google-antigravity) work the same way — npx @mariozechner/pi-ai login <provider>, then set PI_PROVIDER accordingly.

Run

npm install
npm run dev        # watch mode, or:
npm start

Smoke-test:

RELAY_URL=http://127.0.0.1:8788 npm run test:live

Sends a streaming request and prints content + reasoning bytes + tool-call deltas + usage.

Env vars

Var	Default	Purpose
PORT	8788	Listen port (codex proxy uses 8787 — different by default so they coexist)
HOST	127.0.0.1	Listen host. Set to 0.0.0.0 only on a trusted private network
PI_PROVIDER	openai-codex	pi-ai provider id (e.g. anthropic, openai, google)
PI_OAUTH_PROVIDER	same as PI_PROVIDER	Override if the OAuth provider id differs from the model provider id
PI_DEFAULT_MODEL	gpt-5.4	Default model id when the request doesn't pin one
PI_DEFAULT_REASONING	medium	One of minimal, low, medium, high, xhigh
PI_AUTH_PATH	./auth.json	Path to the credentials file written by pi-ai login

Endpoints

• GET /health — liveness + reports the active provider/model/reasoning
• POST /v1/responses — Responses API in, Responses API SSE out

Per-request overrides (headers)

• x-pi-reasoning: minimal | low | medium | high | xhigh — overrides PI_DEFAULT_REASONING (also overridden by the request body's reasoning.effort)
• x-pi-model: <model-id> — overrides the request body's model field
• x-pi-session-id: <id> — passed to pi-ai as sessionId for prompt caching (pi-ai applies this for providers that support it, e.g. OpenAI Codex)

There is no auth on /v1/*. The proxy relies on the network it's deployed on — locally that's loopback; in production that's a Cloudflare VPC subnet reachable only via a Workers VPC connector. Do not expose this on a public interface.

Wire format notes

• Reasoning flows as response.output_item.added (type: reasoning) → response.reasoning_summary_text.delta → response.output_item.done. AI SDK surfaces this as reasoning-start / reasoning-delta / reasoning-end parts.
• Text flows as response.output_item.added (type: message) → response.output_text.delta → response.output_item.done.
• Tool calls flow as response.output_item.added (type: function_call) → response.function_call_arguments.delta → response.output_item.done carrying the final argument string.
• Usage is reported on response.completed as input_tokens / output_tokens (with input_tokens_details.cached_tokens when pi-ai surfaces a cache-read count).
• Errors during streaming are emitted as a response.failed frame followed by data: [DONE] so the client doesn't hang.

Layout

aisdk-pi-proxy/
├── package.json
├── tsconfig.json
├── README.md
├── auth.json              # created by `pi-ai login`, gitignored
└── src/
    ├── auth.ts                  # reads/refreshes auth.json via pi-ai/oauth
    ├── translate-request.ts     # Responses API → pi-ai Context + Tool[]
    ├── translate-response.ts    # pi-ai event stream → Responses API SSE
    └── index.ts                 # Hono server

Self-contained — safe to git mv out to its own repo later.

Deploy

Run on a host inside a Cloudflare VPC subnet. Bind to the private interface only — no public ingress. The Worker reaches it via a Workers VPC connector binding; the connector is the only network path, so no auth lives on the proxy itself. Same operational model as aisdk-codex-proxy.

Relationship to aisdk-codex-proxy

Once this proxy is verified end-to-end against ChatGPT Plus OAuth, it can fully replace ../aisdk-codex-proxy/. The Codex auth file layouts differ — this proxy uses pi-ai's auth.json (run pi-ai login openai-codex), not Codex CLI's ~/.codex/auth.json. Until the swap is confirmed, both can coexist on different ports (8787 codex, 8788 pi).