JetBot

Browser-based AI Coding Assistant — Zero Install. Zero Deploy. Zero Config.

中文文档

JetBot is a fully browser-based AI coding assistant. No backend server, no deployment pipeline, no installation required — just open the page and start working. All code executes client-side via JavaScript; LLM calls go directly from the browser's fetch() to your chosen API provider.

┌─────────────────────────────────────────────┐
│                  Browser                     │
│                                              │
│  ┌──────────┐  ┌───────────┐  ┌───────────┐ │
│  │  React   │  │  Agentic  │  │  Virtual   │ │
│  │   UI     │←→│   Loop    │←→│   FS       │ │
│  └──────────┘  └─────┬─────┘  │ (IndexedDB)│ │
│                      │        └───────────┘ │
│                      │ fetch()               │
└──────────────────────┼──────────────────────┘
                       ↓
              ┌────────────────┐
              │  LLM Provider  │
              │ OpenAI/DeepSeek│
              │ Ollama/Custom  │
              └────────────────┘

Quick Start

git clone https://github.com/hongxin/JetBot.git
cd JetBot/jetbot
npm install
npm run dev

Open http://localhost:5173, select an LLM provider, enter your API key, and start chatting.

Local Ollama (fully offline):

ollama serve
ollama pull qwen3.5:27b
# Select the Ollama provider in JetBot — no API key needed

Production build:

npm run build
# Outputs dist/ — pure static files deployable to any static host

Features

Agentic Loop

An autonomous tool-calling loop where the LLM selects tools, executes actions, and analyzes results until the task is complete. Supports up to 100 iterations with circuit breaker protection (3 consecutive failures) and automatic duplicate-error detection.

10 Built-in Tools

Tool	Permission	Description
read_file	safe	Read files from the virtual file system
list_dir	safe	List directory structure
search_text	safe	Regex search across file contents
write_file	risky	Create or overwrite files
edit_file	risky	Precise find-and-replace editing
http_get	risky	HTTP requests (automatic CORS proxy fallback)
js_eval	risky	Sandboxed JavaScript execution (10s timeout)
render_html	risky	Render HTML/CSS to the preview panel
shell_execute	dangerous	Sandboxed shell commands
export_file	safe	Download files from VirtualFS to local disk

18 Built-in Skills

Activate via /skill <name> to inject domain expertise into the system prompt:

debug · code-review · architect · explain · tdd · writing · refactor · visualize · decision · security · perf

ZPower (Five Elements): zpower · z-observe · z-design · z-build · z-verify · z-evolve · z-diagram

Multi-Provider LLM

Provider	Base URL	Notes
OpenAI	api.openai.com/v1	GPT-4o, etc.
DeepSeek	api.deepseek.com/v1	Cost-effective
Ollama	localhost:11434/v1	Local, no API key
Custom	User-defined	Any OpenAI-compatible endpoint

Cosmos View

Conversations rendered as a force-directed starfield canvas — and beyond. Five node kinds coexist on a single graph:

• user / assistant / tool — the conversation flow
• memory (teal droplet) — facts/decisions saved to the persistent MemoryStore
• skill (amber hexagon) — capabilities distilled from a session and crystallized into SKILL.md

Memory and skill nodes attach via dashed derives edges back to the assistant turn that produced them, making it visible at a glance "what this conversation left behind." Node detail cards are resizable on eight handles and scroll long markdown internally. Drag-connect any two nodes to trigger an LLM relationship analysis.

Permission System

Three-tier access control: safe (auto-approve), risky (approve once, then remembered), dangerous (approve every time). Toggle /auto on for autonomous mode.

Additional Capabilities

• Session Persistence — Every turn is incrementally written to IndexedDB; crashed sessions can be recovered; previous sessions are searchable and recallable as long-term context (/sessions)
• Memory Store — Persistent facts, preferences, decisions injected into the system prompt across sessions (/memory add|list|remove|clear)
• Skill Lifecycle — Distilled skills carry quality score, use count, and a stage (new / active / stable / stale / deprecated); a Skills Panel in Settings manages activation, import (paste SKILL.md), and export
• Auto-Distillation — After a tool-heavy turn with good success rate, the agent proposes a reusable skill via a DistillCard; one click saves it
• File Bridge — Drag-and-drop local file import; export_file triggers browser download
• Scheduler — In-browser scheduled tasks (interval / cron / once), persisted in IndexedDB
• Bilingual i18n — English and Chinese UI, switchable in settings
• Runtime Detection — Auto-detects browser capabilities; tools load dynamically based on what's available

Commands

Command	Description
/help	Show available commands
/clear	Clear conversation history
/status	Display model, tokens, and run status
/model	Show current model
/runtime	Show runtime environment and capabilities
/plan <goal>	Enter or exit plan mode
/next	Advance to next plan phase
/skill <name>|list|off|status|delete|export	Manage skills (activate, list, lifecycle status, delete, export SKILL.md)
/memory list|add|remove|clear	Manage persistent memory entries
/sessions list|search|recall|prune	Browse archived sessions, search across them, prune by age
/export <path>	Download a file from VirtualFS
/schedule	Manage scheduled tasks
/auto on|off	Toggle autonomous mode

Documentation

• Architecture — Source structure and tech stack
• Changelog — Development history and milestones
• Design — Original design document and feasibility analysis
• Idea — Project origin

Chinese versions available: architecture · changelog · design · idea

Design Philosophy

Dao-Qi Unity — Harness AI tools with classical wisdom; think and act in concert, human and machine as one.

• Simplicity — The simplest solution that works. The browser is the Agent.
• Adaptability — Embrace change. Modular architecture evolves continuously.
• Integrity — Hold the line on quality. Parameter validation, circuit breakers, duplicate detection.

Author

Hongxin Zhang — github.com/hongxin

License

MIT