A Claude Science rival that runs on pocket change

What it does OpenAI4S is a hybrid research agent with a split personality: JSON tool calls handle orchestration, permissions, and metadata, while persistent Python and R kernels execute the actual science—computations, simulations, and data analysis. It ships with 604 bundled Skills, connects to seven public scientific databases, and keeps an append-only ledger of every action and artifact. The stack runs locally under hardened sandboxing and can publish read-only session snapshots through an outbound relay you control.

The interesting bit The entire engine and web server are built on Python’s stdlib alone—http.server, a hand-rolled WebSocket, and urllib for LLM traffic—so there is no framework dependency in the core. That austerity is paired with a deliberate two-plane architecture where code cells, not chat loops, do the heavy lifting; a single Python cell can filter, sort, and plot a 100k-row DataFrame without the fourteen ReAct round-trips a tool-only agent would need.

Key highlights

• Runs on the ¥9.9/month Doubao “Small” plan via Volcengine Ark, with one-line UI switches to OpenAI, Anthropic, Gemini, GLM, Kimi, DeepSeek, and MiniMax.
• Pure stdlib core: the engine and web server use only http.server, a hand-rolled WebSocket, and urllib—no FastAPI, no Flask, no external HTTP client.
• Persistent Python and R kernels with synchronous mid-cell host RPC, versioned artifacts, and object-level data lineage.
• Hardened local execution via macOS Seatbelt or Linux bubblewrap sandbox adapters, plus durable human approvals and generation-bound host.bash capabilities.
• 604 bundled Skills—43 curated for GPU and model science plus 561 pinned MIT-licensed bioSkills—installed as code recipes rather than JSON schemas.

Caveats

• v0.3.0 ships the first Windows/WSL2 package but no macOS disk image, so Mac users must install from PyPI.
• General remote compute via SSH and NVIDIA NIM is explicitly labeled a Prototype surface.
• Upgrading from v0.2.x requires care because the database schema jumped from 27 to 32 tables.

Verdict Researchers who want a local, auditable science agent with persistent kernels and a ¥9.9 API bill should look here; developers seeking a simple chat wrapper or a mature remote-compute layer may find the prototype surfaces and stdlib asceticism either refreshing or constraining.