# Eisoptric > Eisoptric is a small, independent AI research project founded and run by Rick Weng in Taiwan. It studies how capabilities form during pretraining, what trained models compute inside (interpretability), and how to keep models that act through tools under human oversight. It releases the tools it builds for that work: Eisoptric Workbench, an open-source desktop research workbench, and Eisoptric MCP, a gateway that puts every tool call an AI agent makes under policy, human approval and audit. - Website: https://eisoptric.com/ - Contact: hello@eisoptric.com - GitHub: https://github.com/Eisoptric - Research status: research notes are in progress; none are published yet. - Eisoptric MCP console: https://console.eisoptric.com - Eisoptric Workbench source: https://github.com/Eisoptric/Workbench - Eisoptric Workbench releases: https://github.com/Eisoptric/Workbench/releases - Sponsorship: https://github.com/sponsors/Eisoptric - Price: both tools are free, with every feature on every account. There are no paid plans. - License: Eisoptric Workbench is licensed under the Apache License 2.0. - Operator: Rick Weng, an individual in Taiwan, who operates the website, the hosted gateway and Eisoptric Workbench distribution. Each link below is a Markdown copy of a page, generated from the same source as the page itself. ## English - [Eisoptric — Research on how AI models learn, and what they compute](https://eisoptric.com/index.md): Eisoptric is a small, independent AI research project founded by Rick Weng in Taiwan. We study how capabilities form during pretraining and what trained models compute inside, and we release the tools we build along the way: Eisoptric Workbench and Eisoptric MCP. - [Eisoptric Research — Pretraining dynamics, interpretability and agent oversight](https://eisoptric.com/research.md): Eisoptric’s research agenda: how capabilities form during pretraining, what trained models compute inside, and how to keep models that act through tools under human oversight. Research notes are published here. - [Lilith — Small open models for studying how models learn](https://eisoptric.com/lilith.md): Lilith is Eisoptric’s series of small language models, trained from scratch so that every checkpoint, the data order and the code can be studied. The Lilim are its smaller siblings. No model has been released yet. - [About Eisoptric — A small, independent AI research project](https://eisoptric.com/about.md): Eisoptric is a small AI research project run by Rick Weng. Who we are, what we work towards, why we build and release tools, and how to reach us. - [Eisoptric MCP — The MCP gateway for AI agents](https://eisoptric.com/mcp.md): Eisoptric MCP is the MCP server your AI agents talk to. It exposes only approved tools, checks every call against policy, and records the audit trail. - [Eisoptric Workbench — Free, open-source software for AI research](https://eisoptric.com/app.md): Eisoptric Workbench is a free, Apache 2.0-licensed desktop workbench for AI research. Run experiments on your own computer or an SSH host, read every sample, look inside the model, and choose how much the assistant may do. - [Eisoptric Trust — How keys, payloads, and credentials are protected](https://eisoptric.com/trust.md): Eisoptric uses zero-access encryption (AES-256-GCM and ML-KEM-1024) for stored audit payloads. Passwords are hashed in the browser before they reach the gateway. Eisoptric cannot unwrap member private keys. - [Eisoptric Pricing — Free Workbench and MCP Gateway](https://eisoptric.com/pricing.md): Eisoptric is free. The desktop App and the MCP gateway cost nothing, and every account gets every feature. Sponsors keep it that way. - [Terms of Service — Eisoptric](https://eisoptric.com/terms.md): The agreement you accept by using Eisoptric. The services are provided by Rick Weng, who also owns the intellectual property. - [Privacy Policy — Eisoptric](https://eisoptric.com/privacy.md): How Rick Weng handles personal data when you use the Eisoptric website, console, MCP backend, Eisoptric Workbench, and Eisoptric Workbench backend. ## 中文(繁體) - [Eisoptric — 研究 AI 模型如何學習、內部在算什麼](https://eisoptric.com/zh/index.md): Eisoptric 是 Rick Weng 在台灣創立的小型獨立 AI 研究計劃。我們研究能力如何在預訓練中形成、訓練好的模型內部在計算什麼,並把研究途中打造的工具釋出:Eisoptric Workbench 與 Eisoptric MCP。 - [Eisoptric 研究 — 預訓練動態、可解釋性與 Agent 監督](https://eisoptric.com/zh/research.md): Eisoptric 的研究方向:能力如何在預訓練中形成、訓練好的模型內部在計算什麼,以及如何讓透過工具行動的模型留在人的監督之下。研究筆記會發表在這裡。 - [Lilith — 用來研究模型如何學習的小型開放模型](https://eisoptric.com/zh/lilith.md): Lilith 是 Eisoptric 從頭訓練的小型語言模型系列,每個 checkpoint、資料順序與程式碼都會公開供人研究。Lilim 是它較小的手足。目前尚未釋出任何模型。 - [關於 Eisoptric — 小型獨立 AI 研究計劃](https://eisoptric.com/zh/about.md): Eisoptric 是由 Rick Weng 經營的小型 AI 研究計劃。我們是誰、為了什麼而做、為什麼打造並釋出工具,以及如何聯絡我們。 - [Eisoptric MCP — AI Agent 的 MCP Gateway](https://eisoptric.com/zh/mcp.md): Eisoptric MCP 是 Agent 實際連線的 MCP Server。它只露出核准過的工具,對每一筆呼叫比對政策,並留下審計軌跡。 - [Eisoptric Workbench — 免費、開源的 AI 研究工作台](https://eisoptric.com/zh/app.md): Eisoptric Workbench 是免費、Apache 2.0 授權的桌面 AI 研究工作台。在自己的電腦或 SSH 主機上跑實驗、讀每一個樣本、檢視模型內部,並決定研究助理可以做多少。 - [信任 — 金鑰、Payload 與憑證如何被保護](https://eisoptric.com/zh/trust.md): Eisoptric 對已儲存的稽核內容採用零存取加密(AES-256-GCM 與 ML-KEM-1024)。密碼在送到 Gateway 之前先於瀏覽器雜湊。Eisoptric 無法解開成員私鑰。 - [方案與定價 — 免費 Eisoptric Workbench 與 MCP Gateway](https://eisoptric.com/zh/pricing.md): Eisoptric 完全免費。Workbench 與 MCP Gateway 都不收費,每個帳號都能使用全部功能,由贊助者支持維持。 - [服務條款 — Eisoptric](https://eisoptric.com/zh/terms.md): 你使用 Eisoptric 時同意的契約。本服務由 Rick Weng 提供,智慧財產也由 Rick Weng 所有。 - [隱私權政策 — Eisoptric](https://eisoptric.com/zh/privacy.md): 你使用 Eisoptric 網站、主控台、MCP 後端、Eisoptric Workbench 與 Eisoptric Workbench 後端時,Rick Weng 如何處理個人資料。 ## Optional - [Full text](https://eisoptric.com/llms-full.txt): Every page above in one file, English first, then 中文. - [Sitemap](https://eisoptric.com/sitemap.xml): Every public page in English and 繁體中文.