SinoBee searches the web, reads files, writes code, runs commands and operates the browser, mouse and keyboard; it also runs scheduled background work and produces complete videos. It delivers tables, charts, files and films. Risky actions ask you first, you can add instructions or stop it mid-task, and it remembers your preferences and gets better over time.
Use online is coming soon. Until then, download SinoBee for free and run it on your own computer.
git clone https://github.com/CAFECA-IO/SinoBee.git && cd SinoBee
docker compose up -d --buildFrom one-off research to long-term tracking, from operating your computer to building a whole project: say it in one sentence and get the finished work back.
Reads web pages, news, documents and GitHub projects, analyzes CSV, Excel and database exports, cross-checks sources and delivers cited reports, tables and charts.
e.g. "Compare GPU prices across three cloud providers, with sources"
Name topics and items to watch: it gathers intelligence on a schedule, tags what changed and assesses the impact, comparing each run with the last; when you need depth, a briefing arrives within 2 hours.
e.g. "Every morning, track market news for every Sino Stock Saver product"
Drives a real browser and desktop apps, and works with files, databases and remote hosts: signing in, filling forms, downloading reports, moving data. Docker mode adds a virtual desktop you can watch and take over; risky actions always ask first.
e.g. "Sign in to the admin panel and download last month’s orders"
Reads links, documents or folders and turns the know-how into reusable skills; your feedback and every task’s experience are folded into memory, so its knowledge base keeps growing.
e.g. "/learn this internal handbook"
Splits big jobs into milestones run by parallel sub-agents; writes code, runs tests and uses git, with independent acceptance checks. Interrupted work resumes where it stopped, and a spent budget still delivers what is done.
e.g. "Port this Flask service to FastAPI with all tests passing"
One sentence to a short film, documentary or ad: script, characters, storyboard, voice-over, rendering, editing and subtitles, with a draft to approve before the final render and an asset kit that keeps sequels consistent.
e.g. "Make a 30-second documentary about Taiwanese tea"
Big jobs done in stages, thinking and doing split between models, a memory of you and of its mistakes, and brakes and records at every step.
Independent milestones go to sub-agents in parallel, and each one starts from a clean context. Progress, not a step count, decides when to stop; interrupted work resumes where it left off, and if the budget runs out you still get what is finished.
A thinking model plans, advises when work is stuck and checks the result before delivery; an execution model does every step; a reviewer looks at the actual files and screenshots with a fresh context. When the job cannot be done, it says so and explains why.
Memory has four layers: task, short-term, long-term and permanent. What gets used stays, what doesn't fades, and what you ask it to keep forever is never changed. One thumbs-down turns a mistake into a check that runs before every delivery.
It asks before deleting files, installing software or force-pushing; skills carry one of 8 trust levels; shares hide keys and paths automatically. What each step did, why it stopped and what it cost are all logged.
What it thought, what it did and what failed, laid out under the answer — and the same record makes the next task better.
✕A conclusion, with no why
✓Every part has a question, an answer and a source
The thinking model diagnoses and sets the next step
Cost, failures and causes are all recorded
A thumbs-down becomes a pre-delivery check
Submit, download, rate and discuss skills, including official ones for AI video production, memory management and database analysis. Upload complete bundles with scripts, references and fixed workflows; reviewed versions are signed with ed25519 and verified by SinoBee on install.
Browse skills →Describe what you want AI to do; AI writes the spec and estimates the cost. People back it at NT$30 a share, and once funded the platform runs it in an isolated container and reports to the backers.
Browse crowdfunds →Anyone can crowdfund an evaluation of a model or a skill. Once the backers accept it, the report is public: time, proposer, skill, model, method, process and result.
See the leaderboard →Build crafted, highly detailed, interactive isometric SVG figures that explain a technical idea as a real physical object (a test rig, a dial gauge, a cabinet of drawers, a plotter, a lock), with many small precise solid parts, four-tone shading, correct painter's order, a framed card with title, hint and live readout, and calm eased motion. Use this whenever someone wants an isometric illustration, an explanatory diagram or figure for docs, a blog post, a landing page or an article, a "Linear-style" or "premium" technical drawing, an interactive explainer, or an SVG that shows how a system, algorithm or component works, even if they never say "isometric". Works framework-free (TypeScript/JS that writes SVG and HTML) and in React/Next.js, and can add realistic WebGL shaders (fire, exhaust, water, caustics, lamp light) that live inside the drawing in the same isometric perspective. Also publishes a finished figure to the Anatomy community gallery when asked.
Researches a question on the web, cross-checks at least two sources and reports findings with citations. Use when the user asks to look up, search, compare or verify facts, prices, news or documentation (查詢、搜尋、比較、調查).
Converts books and documents (PDF, EPUB, DOCX, HTML, Markdown, plain text, RTF, MOBI/AZW with Calibre) into structured agent skills, extracting frameworks, mental models, principles, techniques, and anti-patterns. Use when the user wants to study a document through GitHub Copilot CLI, Amp, Claude Code, Hermes Agent, OpenCode, or OpenClaw, apply an author's frameworks while working, or build a reusable knowledge base from a file.
Analyses whether a technology shown in a video or described in text can work - extracts every testable claim, checks it against physics with a numeric model (required vs limit), gives a verdict (feasible, feasible within limits, infeasible) and explains the engineering and the physical limits. Then builds an interactive 3D / isometric anatomy page with the anatomy skill - how the system works and where it stops if feasible, or why each key property cannot be achieved if not - plus a Markdown report. Use for 技術可行性、工程原理、物理極限、這可能嗎、真的做得到嗎、科技產品或發表會影片分析, feasibility study, is this physically possible, how does it work, debunk or explain a tech demo.
Watches a YouTube video (metadata, full transcript with timestamps, and stills about every 10 seconds) and writes a Markdown report with images - a SinoBee.ai-style cover, what the video says, a timestamped outline, an analysis section chosen from 16 video categories, and a fact-check of its claims for news, politics, science, finance, health, tech and documentary videos. Use when the user gives a YouTube link and asks to summarise, analyse, review, fact-check or report on it (YouTube 影片分析、影片摘要、幫我看這部影片、查核影片內容).
Manages the agent's layered memory - distills the key points of a task into task, short-term, long-term or permanent memories, compresses memory when it grows, and adds or removes memories on the user's instruction. Use when the user asks to remember, forget, pin, review, clean up or compress memories (記住、永久記住、忘記、刪除記憶、整理記憶、壓縮記憶、你記得什麼), or at the end of a long task to keep what matters.