A local AI coding agent in TypeScript — agentic loop, streaming, 9-layer permissions, context compaction, memory, file undo. Inspired by Claude Code. Supports DeepSeek / OpenAI / Anthropic.
A fully local, terminal-based AI coding agent built in TypeScript.
Architecture deeply inspired by Claude Code — agentic loop, tool system, permission model, context management, and more.
Supports DeepSeek · OpenAI · Anthropic — swap with one env var.
| Category | What it does |
|---|---|
| Agentic loop | Autonomous multi-turn while loop: AI decides the next tool call, executes it, feeds results back, repeats until done |
| Real-time streaming | Token-by-token output for both Anthropic and OpenAI-compatible APIs; tool-call JSON fragments accumulated across chunks |
| 9 built-in tools | ReadFile · WriteFile · StrReplace · Bash · Glob · Grep · Git · WebFetch · SpawnAgent |
| Permission system | 9-layer decision chain: hardcoded bans → always-allow → session cache → .assistant-rules.json → mode → headless guard → read-only fast-path → user prompt |
| Bash risk analysis | Pre-execution command analysis: strips timeout/sudo wrappers, classifies risk (low/medium/high/critical), color-coded prompt |
| Context compaction | Round-based grouping → LLM summarises older rounds when token count exceeds threshold; keeps recent N rounds intact |
| Persistent memory | Per-project MEMORY.md injected into system prompt each session; auto-extracted and saved on exit |
| File history & undo | SHA256-named snapshots before every write; /undo restores all files changed in the last turn |
| Multi-agent | SpawnAgent tool spawns isolated sub-agents; --coordinator mode decomposes tasks into Research→Synthesis→Implementation→Verification with parallel dispatch |
| Prompt cache | Injects cache_control: ephemeral on system prompt + last user message (Anthropic) — ~90 % cost reduction on repeated prefixes |
| Rate-limit retry | Exponential backoff + jitter + Retry-After header parsing for 429/503 errors |
| *Hooks system |
HAL 分层混合模型工作流 — 强模型(Claude)负责理解/拆解/验收,低成本模型(DeepSeek)负责检索/提取/清洗。Hermes Agent skill。
An LLM agent fine-tuned on DeepSeek for spaced repetition, dynamically integrating knowledge points based on the Ebbinghaus forgetting curve.
基于 STM32F103 构建的端到端 AI 智能手表生态。自研“零重定位”原生机器码动态加载引擎与页面栈式 UI 框架;集成生产级 OTA 回滚保护机制与高带宽(921600 baud)串口协议栈。通过 Node.js 中继实现 DeepSeek AI 语义控制及 ASRPRO 语音全双工交互,是一个集成了分布式计算、现代存储管理与 AI Agent 的嵌入式全栈工程。
A Meta-Agent-Driven Self-Evolving Multi-Agent System for UAV Detection and Tracking
One command to run Hermes AI Agent with a browser UI. Zero prerequisites. 一行命令,AI 就位。
网页应用Agent,接入DeepSeek、Mimo等模型
Workflows from the Neura Market marketplace related to this DeepSeek resource