Lele

Lightweight personal AI assistant in Go — single binary, small footprint, fast TUI.

中文 | 日本語 | Português | Tiếng Việt | Français | Español | English

---

Lele is a workspace-first AI assistant built for long-running hosts: CLI and TUI chat, a multi-channel gateway, web UI, native client API, skills, cron, and subagents — without a JS runtime or multi-process TUI stack.

Quick Start

# Linux / macOS / BSD
curl -fsSL https://raw.githubusercontent.com/xilistudios/lele/main/install.sh | sh

# Windows (PowerShell)
irm https://raw.githubusercontent.com/xilistudios/lele/main/install.ps1 | iex

# From source
git clone https://github.com/xilistudios/lele.git && cd lele
make deps && make build && make install

# Initial setup
lele onboard
# CLI agent interface
lele agent -m "What can you do?"
# TUI
lele tui

Installers verify SHA256 and install to ~/.local/bin by default.

Benchmarks

Measured on Apple M1 / macOS arm64. Idle TUI sessions under a PTY, no model calls. Render is wall time to paint one full frame (lele: in-repo View() bench at 200×50; others: time from process start to first complete TUI paint — init + first frame). RSS is peak process-tree memory ~6s after that paint.

Tool	Runtime	Install size	Render	Idle TUI RSS	Processes
lele	Go + Bubble Tea	57 MB single binary	≈ 3.4 ms/frame (200×50 View())	~42 MB	1
Claude Code 2.1.267	Bun-compiled TS	191 MB binary	~400 ms to first paint	170–187 MB	1
pi 0.85.1	Bun-compiled TS + pi-tui	71 MB binary	~350 ms to first paint	~173 MB	1
OpenCode 1.17.8	Bun + OpenTUI	123 MB binary	~950 ms to first paint	~1.0 GB peak	1
Hermes 0.14.0	Python + Ink	~1 GB tree	~1.3 s to first paint	~374 MB	3–4

Resource difference vs others

Same machine, same idle TUI session. Rough multipliers vs lele (~42 MB RSS, ~57 MB binary, ~60 ms to first TUI paint):

vs	Memory	Disk / install	First render	Notes
Claude Code	~4× less	~3× smaller	~7× faster	42 vs ~180 MB · 60 ms vs ~400 ms
pi	~4× less	~equal	~6× faster	42 vs ~173 MB · 60 ms vs ~350 ms
OpenCode	~25× less	~2× smaller	~16× faster	42 vs ~1 GB · 60 ms vs ~950 ms
Hermes	~9× less	~17× smaller	~22× faster	42 vs ~374 MB · 60 ms vs ~1.3 s

Features

Agent runtime

• Tool-using agent loop with iteration limits
• MCP client: connect to stdio/HTTP/SSE MCP servers through layered mcp.json — see docs/mcp.md
• Named agents, model fallbacks, per-agent thinking_level (override with /think)
• Session persistence (SQLite) and optional ephemeral sessions
• File attachments in native/web flows

Interfaces

• CLI (lele agent) and full Bubble Tea TUI (lele tui)
• Built-in web UI + native REST/WebSocket client with PIN pairing
• Gateway for chat channels (Telegram, Discord, Slack, WhatsApp, Feishu, Line, QQ, DingTalk, …)

Automation

• Scheduled jobs (lele cron) and HEARTBEAT.md heartbeat tasks
• Skills and async subagents
• Multi-agent group chat (MoA / round-robin / moderator / pipeline) — see docs/moa-group-chat.md

Safety

• Workspace restriction, exec deny patterns, approval flow
• Token auth for native clients, upload limits/TTL

Screenshots

TUI

lele tui              # new session
lele tui -s <id>      # resume session

Streaming chat, tool visualization, markdown, session sidebar, context stats. Themes: dracula (default), nord, catppuccin, gruvbox, tokyo-night, solarized-light — switch in Settings → Interface, stored in ~/.lele/tui.json. Language via LELE_LANG or /lang.

TAB cycles your configured agents in place; /mode [agent|chat|group] switches the session filter (bare /mode advances to the next one).

Web UI

1. lele onboard — enable web UI, get a pairing PIN
2. lele gateway — serves /, /api/v1/*, WebSocket on port 18790
3. Open the browser and pair with the PIN

Full client API: docs/client-api.md.

Configuration

Config lives at ~/.lele/config.json (template: config/config.example.json).

{
  "agents": {
    "defaults": {
      "workspace": "~/.lele/workspace",
      "restrict_to_workspace": true,
      "model": "glm-4.7",
      "max_tokens": 8192,
      "max_tool_iterations": 20
    }
  },
  "providers": {
    "openrouter": {
      "type": "openrouter",
      "api_key": "YOUR_API_KEY"
    }
  }
}

Main sections: agents.defaults, session, channels, providers, tools, heartbeat, gateway, logs, devices.

Providers: anthropic, openai, openrouter, groq, zhipu, gemini, vllm, nvidia, ollama, moonshot, deepseek, github_copilot, plus hermes-agent ports (xai, nous, lmstudio, minimax, vercel, opencode, opencode_go, huggingface, novita, xiaomi, …), Qwen Cloud / Alibaba Token Plan (alibaba_token_plan, alibaba_token_plan_cn), and named OpenAI-compatible backends (model, context_window, vision, reasoning, …). A curated model catalog (catalog/ in-repo, cached under ~/.lele/cache/catalog/) with context/vision/thinking metadata powers WebUI/TUI autocomplete when adding models.

Workspace (~/.lele/workspace/): sessions/, memory/, state/, cron/, skills/, AGENT.md, HEARTBEAT.md, IDENTITY.md, SOUL.md, USER.md.

CLI

Command	Description
lele onboard	Initialize config and workspace
lele agent / lele agent -m "..."	Interactive or one-shot agent
lele tui / lele tui -s <session>	Terminal UI
lele gateway	Messaging gateway + web UI
lele auth login	Authenticate providers
lele status	Runtime status
lele cron list / lele cron add ...	Scheduled jobs
lele skills list	Installed skills
lele client pin / lele client list	Native client pairing
lele version	Version info

Development

Requires Go 1.25+, Bun (web UI), and Make.

make deps web-build build   # build binary + web
make test fmt vet           # checks
make build-all              # cross-compile

License

MIT