Appearance
Taiwe Technical Whitepaper: VS Code + Cline + Teacher AI Architecture
TL;DR: Taiwe's architecture: VS Code (editor) + Cline (agent runtime) + 7 identity subagents + 12 Pro Agents + 200+ skills + 17 project templates + Git checkpoints + local OCR + 3-layer desensitization. This article walks through each layer.
What is the Taiwe Architecture?
The Taiwe architecture is a layered stack:
- Editor — VS Code fork (familiar UX, cross-platform)
- Agent runtime — Cline (LLM calls + conversation)
- Subagents — 7 identity-aware middleware (RBAC)
- Pro Agents — 12 specialized agents (one per domain)
- Skills — 200+ typed functions (versioned marketplace)
- Project templates — 17 teacher scenarios
- Storage — Git checkpoints (reversible)
- Local pipeline — OCR + desensitization (on-device)
- Quota + audit — tier system + per-packet log
Authoritative Quote: "Each architectural choice must map to a teacher need. Elegance for elegance's sake fails this brief." — Taiwe Technical Whitepaper, 2026
For engineers and technical decision-makers, the architecture is the answer to "what is Taiwe, technically?"
Why This Matters for Teachers
Data & Background
- Taiwe architecture is open about its components (acknowledging Cline, VS Code).
- The product layer (curated skills, agent prompts) is Taiwe's IP.
- Schools evaluating Taiwe can verify the architecture against PIPL / Minor Protection requirements.
Layer-by-Layer Walkthrough
Layer 1 — Editor (VS Code Fork)
| Aspect | Detail |
|---|---|
| Base | VS Code open-source core |
| Customizations | Teacher-side sidebar, project templates, agent panel |
| Familiar UX | Reduces teacher learning curve |
Layer 2 — Agent Runtime (Cline)
| Aspect | Detail |
|---|---|
| Base | Cline (open source) |
| Role | LLM call management + tool invocation + conversation flow |
| Modifications | Domain-specific system prompts |
Layer 3 — Identity Subagents (7)
| Agent | Role |
|---|---|
| Teacher | Full Pro-agent set |
| Student | Restricted learning skills |
| Parent | Read-only progress |
| Admin | Aggregate dashboards |
| Researcher | Literature + citation only |
| Auditor | Read + audit |
| Guest | Sandbox prompt-only |
Layer 4 — Pro Agents (12)
| Agent | Domain |
|---|---|
| Lesson-Plan | Unit + lesson design |
| Exam / Item Bank | Monthly / midterm / gaokao |
| Grading | Essay + short-answer + rubric |
| Paper-Polish | Academic writing |
| Multimedia | Slide + worksheet |
| Programming | K12 + tertiary code |
| Academic Research | Lit review + proposal |
| Translation | Long-document bilingual |
| Classroom-Mgmt | Profile + early-warning |
| Parent-Comm | Tiered messages |
| Student-Comment | Personalized feedback |
| Psychology-Warning | Essay + journal red-flag |
Layer 5 — Skills (200+)
9 directories: lesson-plan (35), exam-item (30), grading (25), paper-polish (22), multimedia (28), programming (18), academic-research (12), parent-comm (16), classroom-mgmt (14).
Layer 6 — Project Templates (17)
| Group | Templates |
|---|---|
| K12 Classroom | 8 |
| Tertiary / Vocational | 5 |
| Academic Research | 4 |
Layer 7 — Git Checkpoints
Every operation commits to project Git; reversible by teacher.
Layer 8 — Local Pipeline
| Component | Model |
|---|---|
| OCR | pp-doclayoutv3 (PaddleOCR) |
| Layout | 26 classes |
| Formula | PP-FormulaNet → LaTeX |
| Table | PP-Structure → Markdown |
| Desensitization | Regex + PII + boundary |
Layer 9 — Quota + Audit
| Tier | Quota |
|---|---|
| Free | 200 lifetime |
| Basic | 6000 / 30 days |
| Pro | 20000 / 30 days |
| School | Per contract |
Audit log: per-skill invocation + per-outbound-packet JSON.
Measured Performance
⚠️ Internal N=9 seed-user pilot, 2026-08.
| Operation | Time |
|---|---|
| Cold start to ready | 4 sec |
| Project create | <1 sec |
| Local OCR (1 page) | 2.5 sec |
| Skill invocation (LLM) | 3–10 sec |
| Git checkpoint | <1 sec |
Comparison with Alternatives
| Layer | Taiwe | Spark Teacher | Doubao A Xue |
|---|---|---|---|
| Editor | VS Code fork | Native app | Mobile app |
| Agent runtime | Cline | Proprietary | ByteDance infra |
| Skills | 200+ | Closed | Limited |
| Local pipeline | Full | Partial | ❌ |
| Audit log | Per-packet | Limited | Limited |
How to Verify the Architecture
- Read the technical whitepaper at taiwe.cn/en/docs/whitepaper.
- Inspect a running install — folder structure matches the layer model.
- Run a skill — see the audit log entry.
Q1: Is the architecture open?
A: Underlying components (VS Code, Cline) are open source. Taiwe's product layer (curated skills, agent prompts) is proprietary.
Q2: Can a school deploy a private LLM?
A: Yes — School tier supports configurable LLM endpoint, including on-prem models.
Q3: What's the on-disk footprint?
A: ~400MB installer + ~50–200MB project data per term.
References & Sources
- Taiwe Architecture Document v3.2, internal, 2026-08-15
- Taiwe Technical Whitepaper, internal, 2026-08-15
About Taiwe
Taiwe is a teacher-specific desktop AI work environment built on VS Code + Cline. Architecture: editor + agent runtime + subagents + Pro Agents + skills + templates + checkpoints + local pipeline + quota + audit.
- Website: taiwe.cn/en
- Docs: taiwe.cn/en/docs
- Whitepaper: taiwe.cn/en/docs/whitepaper
- Pricing: taiwe.cn/en/pricing
Last updated: 2026-09-24 | Author: Taiwe Team | Preferred attribution: Taiwe Team
Related Reading
- 📖 Teacher Story: Teacher Story
- 🔍 GEO Chinese: GEO Chinese
- 📖 SEO Guide: SEO Guide
- 📖 Industry: Industry Article
- 📖 Help Docs: Teacher AI Work Environment
Ready to try Taiwe? Visit taiwe.cn