H
August 26, 2026 · View on GitHub
AI coding agent. 27 tools, 11 sub-agent types, DeepSeek-powered.
~5× less RAM & disk than VS Code.
Tier 1 — vs VS Code OSS (Efficiency)
The lightweight baseline. H cuts the overhead of a general-purpose editor platform by focusing strictly on what an agent needs.
| H | VS Code OSS | |
|---|---|---|
| RAM (idle) | ~200–300 MB | ~1–2 GB |
| Disk size | ~200–300 MB | ~1–2 GB |
| Startup | ~2–3 s (cold) | ~5–10 s (cold) |
| Process model | 3–5 processes (Electron + Node server) | 15–30+ processes (extension hosts, LSPs, watchers, shared workers) |
| Focus | Agent-first workspace | Extensible editor platform |
| License | AGPL-3.0 | MIT |
Tier 2 — vs VS Code OSS + Copilot (Scope)
Copilot is a code assistant — autocomplete, chat, and inline suggestions. H is agent-native: it plans, delegates, remembers, and operates the browser and terminal directly.
| H | VS Code OSS + Copilot | |
|---|---|---|
| Model | Agent — plans, executes end-to-end tasks, iterates on its own errors | Assistant — responds to prompts, suggests snippets, requires the user to drive |
| Delegation | delegate_task — spawns 11 specialized sub-agents (browser, code-writer, researcher, security-auditor, architect…) with isolated context windows | None — single chat thread |
| Live sub-agent UI | Per-agent color-coded streaming, progress cards, nested todos | None |
| Persistent memory | File-based cross-session memory (remember / recall / forget) | Per-session chat only |
| Browser automation | Built-in Electron webview + 15 browser tools (pixel screenshots, coordinate-based click/type, upload, console, network) | None |
| Terminal control | run_in_terminal — real PTY with agent read/write/kill access | Integrated terminal, no agent access |
| Knowledge graph | Auto-built codebase graph (.kg): symbol-level imports, stdlib/third-party classification, unused-import detection, read_graph tool (6 query types) | None |
| Multi-agent orchestration | Parent → child agent handoff with streaming summaries | None |
| Step-by-step locking | IDE-enforced todo progression — sub-agents can't skip ahead | None |
| Tool count | 27 built-in primitives | Copilot-only feature set (no browser/terminal/graph) |
| MCP | Built-in MCP server (stdio + SSE) | Extensions only |
H is not an editor and not a chat panel — it's a self-driving workspace. The agent owns the task loop: it reads and writes files, runs commands, opens web pages, spawns specialist sub-agents for subtasks, and tracks progress on a locked todo list that the user can follow but the agent can't shortcut.
Quick Start
npm run install:all
npm run dev
Get a key at platform.deepseek.com.
Agent Tools (27 total)
| Category | Tools |
|---|---|
| Filesystem | read_file, write_file, edit_file, list_files, search_files, grep, create_directory, rename_file, delete_file |
| Terminal | run_command (sandbox), run_in_terminal (real PTY), kill_terminal, read_command_output |
| Browser | browser_navigate, browser_info, browser_screenshot, browser_click, browser_type, browser_clear, browser_select, browser_scroll, browser_press_key, browser_wait, browser_move_mouse, browser_right_click, browser_upload_file, browser_console, browser_request_errors |
| Diagnostics | read_problems, read_graph |
| Control | write_todos, write_summary, task_complete, delegate_task |
| Memory | remember, recall, forget |
Sub-Agent Profiles (11 types)
browser · code-search · code-writer · researcher · planner · frontend-specialist · backend-specialist · security-auditor · architect-analyst · docs-analyst · documentation-writer
Each sub-agent runs in its own isolated context window. The parent agent delegates tasks and streams results live to the UI with per-agent color coding.
Tech Stack
- Client: React 18, Vite, Monaco Editor, xterm.js
- Server: Node.js, Express, WebSocket (ws), node-pty
- AI: DeepSeek API (tool-calling, embeddings, prefix caching)
- Desktop: Electron
- LSP: 30+ language servers (pyright, gopls, rust-analyzer, clangd, jdtls…)
License
AGPL-3.0 — see LICENSE.
Copyright (c) 2026 Luke Yong.
~250 MB RAM, ~250 MB disk. Agent-first. Open source.
ai coding agent · deepseek · electron · react · typescript · multi-agent · mcp · lsp · knowledge graph · browser automation · terminal · sub-agent delegation