Make AI actually work for you — your personal AI assistant.

Generative power and natural-language automation are the computing power of the AI era.
Agenvoy is your productivity infrastructure.

English · 繁體中文

---

What is Agenvoy

Computing power defined what you could build in the computer age. In the AI era, that role belongs to generative capability and natural-language automation — and like computing power, it needs infrastructure you own.

Agenvoy is that infrastructure. A single Go binary on your machine turns Claude, GPT, and Gemini into a unified, always-on productivity layer. Configure it once, talk to it from anywhere — Telegram, Discord, LINE (alpha), terminal TUI, or browser. Same memory, same tools, same skills across every channel.

What it can do

Teach the agent to build its own tools — no programming required. Ask "what's the weather in Taipei?" and if no weather tool exists, the agent searches a public API catalog, picks a free endpoint, scaffolds a Python script with tool.json schema, sandboxes it, and answers — all in one turn. The tool persists; next time you (or any session) asks, it just runs.

How auto-generation works: the system prompt enforces a Capability Gap rule — whenever a request needs live external data and no registered tool covers it, the agent must create one before answering. It cannot shortcut with raw HTTP or inline scripts. The generated tool lives at ~/.config/agenvoy/tools/script/<name>/ with always_allow: true, so future calls skip confirmation. Auth-required APIs route through store_secret + keychain; the API key never touches the LLM context.

Demo · Auto-generate tools	Demo · Skill-based scheduler
Demo · Sub-agents co-work	Demo · Tool install from registry

Out of the box it also:

• Chats anywhere — Telegram inline buttons · Discord select menus / modals · LINE (alpha) · terminal TUI · browser canvas. One daemon, every surface.
• Speaks replies (TTS), transcribes audio / video.
• Schedules itself — say "every weekday 8am push a Hacker News top-stories digest to Telegram" and the agent builds the cron + push pipeline for you.
• Picks the right model per task — Claude for coding, Gemini for video, GPT for research, routed automatically.
• Searches your files semantically — KuraDB indexes your local docs / notes (file → embedding vector); the agent answers from your own knowledge base, not generic training data.
• Remembers across sessions — three-tier memory: recent context + vector similarity (ToriiDB) + full-text archive (SQLite FTS5). Every message is dual-written; nothing is ever lost.
• Publishes and installs custom tools — share AI-built tools across machines through the pkg.agenvoy.com registry; email-verified uploads with downgrade-proof versioning, one-popup install with dependency auto-resolve.
• Self-improves skills after failures — when a skill execution hits tool errors, the agent rewrites the skill definition (fixes wrong tool names, unclear steps, broken sequences) and commits the fix automatically. The skill gets better every time it fails.

One-line install

curl -fsSL https://agenvoy.com/static/scripts/install.sh | bash

Single Go binary at /usr/local/bin/agen. macOS / Linux. No Node, no Python, no Docker.

Running the daemon on a MacBook? sudo pmset -c sleep 0 keeps the system awake while plugged in — prevents the daemon from being suspended on AC power.

How it compares

Compared against the two closest peers — personal AI agent frameworks with daemon + chat-platform integration.

You want…	Agenvoy	OpenClaw	Hermes
One-line install, single binary	✅ Go	❌ pnpm monorepo	❌ pip + docker
Use Claude + GPT + Gemini in one chat	✅ auto-routed by dispatcher	✅ manual switch	✅ manual switch
Native chat buttons / menus / modals	✅ inline keyboard / select / modal	⚠️ reactions / text-based	⚠️ text-based options
Agent builds & saves its own tools	✅ FaaS-sandboxed scripts + APIs	❌	⚠️ skill-only
First-contact verification on Telegram/Discord	✅ 6-digit OTP	⚠️ pairing code (manual approve)	⚠️ pairing code (gateway/pairing.py)
Cross-session push (any session → chat)	✅ send_to_chatbot	❌	⚠️ send_message tool (scope differs)
Native document RAG (file → embedding)	✅ KuraDB in-process (semantic + keyword)	❌ (conversation memory only)	❌ (conversation memory only)

Looking for the full feature-by-feature breakdown? See What makes it different below.

---

For developers

What makes it different

• Dispatcher-based intelligent routing — a dispatcher model routes every task to the best-fit worker (Claude for coding, Gemini for video, GPT for research), instead of forcing one model to do everything.
• Agent that builds and persists its own tools — when a tool is missing the agent writes a script or API into extensions/ and loads it as a native tool on the next run; MCP servers are supported alongside.
• One runtime across every channel — Telegram, Discord, TUI, Web, and cron all attach to the same daemon; sessions, memory, and the tool set are shared, not rebuilt per surface.

Agenvoy vs Mainstream Products: Full Detailed Comparison

1. Overview

	Agenvoy	OpenClaw	Hermes Agent	Claude Code	Codex CLI	Gemini CLI
Language	Go	TypeScript	Python	TypeScript	Rust + TypeScript	TypeScript
License	Apache 2.0	MIT	MIT	Proprietary	Apache 2.0	Apache 2.0
Author	Individual (pardnchiu)	Community	NousResearch	Anthropic	OpenAI	Google
Primary use	Multi-platform AI Agent framework	Multi-platform AI Agent	Multi-platform AI Agent	Terminal coding assistant	Terminal coding assistant	Terminal coding assistant
Architecture	Daemon + TUI + Chat	Daemon + TUI + Chat	Daemon + TUI + Chat	CLI session	CLI session	CLI session

---

2. AI Provider Support

	Agenvoy	OpenClaw	Hermes Agent	Claude Code	Codex CLI	Gemini CLI
Claude	✅	✅	✅	✅ only	❌	❌
OpenAI / GPT	✅	✅	✅	❌	✅ only	❌
Gemini	✅	✅	✅	❌	❌	✅ only
Codex (OpenAI OAuth)	✅	✅	✅	❌	✅	❌
GitHub Copilot	✅	✅	✅	❌	❌	❌
Nvidia NIM	✅	✅	✅	❌	❌	❌
OpenAI-compat	✅	✅ Ollama/LM Studio	✅ OpenRouter 200+	❌	❌	❌
DeepSeek	✅	✅	✅	❌	❌	❌
xAI (Grok)	✅ API key	✅	✅ OAuth + API key	❌	❌	❌
Mistral	❌	✅	⚠️ via OpenRouter (no dedicated)	❌	❌	❌
Dispatcher routing	✅ dedicated dispatcher model	❌	❌	❌	❌	❌

---

3. Runtime & Frontend

	Agenvoy	OpenClaw	Hermes Agent	Claude Code	Codex CLI	Gemini CLI
TUI	✅ bubbletea	✅ openclaw tui	✅ React Ink	✅ ink	✅	✅
CLI	✅	✅	✅	✅	✅	✅
HTTP API / Web UI	✅ gin	✅ dashboard / webchat	✅ Web Dashboard	❌	❌	❌
Daemon mode	✅ native --daemon	✅ systemd/launchd	✅ gateway daemon	❌	❌	❌
Named sessions	✅	⚠️ workspaces / per-agent sessions	✅ session picker	❌	❌	❌

---

4. Chat Platform Integration

	Agenvoy	OpenClaw	Hermes Agent	Claude Code	Codex CLI	Gemini CLI
Telegram	✅ native daemon	✅ native daemon	✅ native daemon	⚠️ Channels MCP (requires active session)	❌	❌
Discord	✅ native daemon	✅ native daemon	✅ native daemon	⚠️ Channels MCP (requires active session)	❌	❌
iMessage	❌	✅ BlueBubbles	✅ BlueBubbles	⚠️ Channels MCP (macOS only)	❌	❌
LINE	⚠️ alpha (linebot branch)	✅	✅	❌	❌	❌
WhatsApp / Slack	❌	✅ 24+ platforms	✅ 24+ platforms	❌	❌	❌
Always-on receiving (no session needed)	✅ daemon	✅	✅	❌	❌	❌
Cross-session send (any session → chat)	✅ send_to_chatbot	❌	⚠️ send_message tool	❌	❌	❌
First-contact verification	✅ 6-digit OTP (crypto/rand)	✅ pairing code (dmPolicy: pairing)	✅ pairing code (gateway/pairing.py)	❌	❌	❌
Native platform UI (buttons / menus / modals)	✅ inline keyboard / select menu / modal	⚠️ reactions / text-based	⚠️ text-based options	❌	❌	❌

Platform layer: Agenvoy's Telegram and Discord integrations are both built on pardnchiu/go-bot, independently maintained and open source. go-bot encapsulates the bot protocol details for both platforms — Agenvoy only implements business logic, while the platform API layer is entirely handled by go-bot.

Key difference: Claude Code Channels requires an active session. OpenClaw and Hermes have daemons but their in-chat confirmations are largely text/reaction-based. Agenvoy uses native platform UI — Telegram inline keyboards and Discord select menus / modals. Additionally, Agenvoy's cross-session send tools let any session type (CLI, TUI, HTTP, scheduled script) push to a specific Telegram/Discord chat — a capability competitors expose only partially (e.g. Hermes' send_message within its own gateway scope).

---

5. Telegram Feature Comparison

Feature	Agenvoy	OpenClaw	Hermes Agent	Claude Code Channels
Send text reply	✅	✅	✅	✅
Send voice (TTS)	✅ Gemini TTS → OGG	✅ ElevenLabs/Hume	✅ Edge TTS/ElevenLabs	❌
Send file attachments	✅ [SEND_FILE:]	✅	✅	❌
Receive user attachments	✅ photo/doc/voice/video	✅	✅	❌
Voice-to-text (STT)	✅ Gemini, 14 formats	✅ Whisper/Gemini	✅ faster-whisper (local)	❌
Tool confirm (interactive)	✅ native inline keyboard	⚠️ text approval prompt	⚠️ text options	❌
ask_user (picker)	✅ native button/modal	⚠️ /models picker	⚠️ text options, up to 4	❌
Format reference (lazy-load tool)	✅ format_chatbot	❌	❌	❌
Scheduler output push	✅	✅	✅	❌
Cross-session push (from any session)	✅ send_to_chatbot	❌	⚠️ send_message tool	❌
Offline receiving (daemon)	✅	✅	✅	❌

---

6. Discord Feature Comparison

Feature	Agenvoy	OpenClaw	Hermes Agent	Claude Code Channels
Send text reply	✅	✅	✅	✅
Send voice (TTS)	✅ Gemini TTS → OGG/OPUS	✅	✅	❌
Send file attachments	✅ batch 10/message	✅	✅	❌
Receive user attachments	✅ photo/doc/voice/video	✅	✅	❌
Tool confirm (interactive)	✅ select menu button	✅ /model picker	⚠️ text options	❌
ask_user (modal)	✅ select/multi-select/modal	⚠️ limited	⚠️ text options	❌
Format reference (lazy-load tool)	✅ format_chatbot	❌	❌	❌
Guild mention guard	✅	✅	✅	❌
Discord Markdown aware	✅ full spec as lazy-load tool	⚠️ partial	⚠️ partial	❌
Character limit aware	✅ 1600 char hard limit in prompt	❌	❌	❌
Cross-session push (from any session)	✅ send_to_chatbot	❌	⚠️ send_message tool	❌

---

7. Scheduler

	Agenvoy	OpenClaw	Hermes Agent	Claude Code	Codex CLI	Gemini CLI
Cron jobs	✅ SKILL.md + cron	✅ built-in	✅ built-in	✅ cloud-assisted cron/task	❌	❌
One-shot tasks	✅	✅ at format	✅ natural language	✅ cloud-assisted	❌	❌
TUI CRUD	✅	✅ openclaw cron	✅ cronjob tool	❌	❌	❌
fsnotify hot-reload	✅	❌	❌	❌	❌	❌
Push output to Telegram/Discord	✅	✅	✅	❌	❌	❌
AI tool management (add/list/remove)	✅	❌	✅ cronjob tool	❌	❌	❌
Local execution (no cloud required)	✅	✅	✅	❌ cloud-dependent	❌	❌

Scheduler layer: Agenvoy's scheduler is built on pardnchiu/go-scheduler, a self-maintained ecosystem package providing cron expression parsing, one-shot tasks, fsnotify hot-reload, and full output routing back to chat platforms.

---

8. Tool Ecosystem

	Agenvoy	OpenClaw	Hermes Agent	Claude Code	Codex CLI	Gemini CLI
MCP support	✅ client	✅ client	✅ client + server	✅ client	❌	✅ client
Custom tools (script-tool-add)	✅ AI-generated	❌	✅ auto-creates skill	❌	❌	❌
API tool discovery (search-api → add)	✅	❌	❌	❌	❌	❌
Tool registry (publish + install across machines)	✅ pkg.agenvoy.com (Cloudflare Worker + R2 + D1, email verification + downgrade guard)	⚠️ ClawHub (skills + plugins)	⚠️ agentskills.io (skills only)	❌	❌	❌
Skill system	✅ SKILL.md lazy-load	✅ SKILL.md 5400+ community	✅ SKILL.md agentskills.io	✅ CLAUDE.md	❌	❌
Skill self-improvement (auto-fix on failure)	✅ trace → rewrite → auto-commit	❌	✅	❌	❌	❌
Format reference as lazy-load tool	✅ format_chatbot	❌	❌	❌	❌	❌
Document RAG (external knowledge base)	✅ KuraDB (in-process vector + semantic/keyword)	❌ (conversation memory only)	❌ (conversation memory only)	❌	❌	❌
Media transcription STT	✅ Gemini, 14 formats	✅ Whisper/Gemini	✅ faster-whisper (local)	❌	❌	❌
TTS voice output	✅ Gemini TTS	✅ ElevenLabs/Hume/MS	✅ Edge TTS/ElevenLabs/OpenAI	❌	❌	❌
Computer use / browser	✅ go-rod + Playwright MCP	✅ Chrome CDP	✅ browser CDP + computer-use (cua-driver)	✅ beta	❌	❌

Tool sandbox architecture: Agenvoy's Python/JavaScript/API custom tool interfaces are built on the pardnchiu/go-faas (Function as a Service) concept. Each AI-generated tool runs as an isolated function unit with its own lifecycle and security boundary. This is the only FaaS-level sandbox design for tool extension among all compared products.

---

9. Memory System

	Agenvoy	OpenClaw	Hermes Agent	Claude Code	Codex CLI	Gemini CLI
Instruction file system	✅ SKILL.md	✅ SKILL.md	✅ SKILL.md	✅ CLAUDE.md	❌	❌
Conversation history search	✅ Three-tier: context + ToriiDB vector + SQLite FTS5	✅ LanceDB vector	✅ SQLite FTS5	❌	❌	❌
External document RAG (native, in-process)	✅ KuraDB (semantic + keyword, OpenAI embeddings)	❌ (use MCP)	❌ (use MCP)	❌	❌	❌
Error memory	✅ ToriiDB	❌	❌	❌	❌	❌
Action log	✅	❌	❌	❌	❌	❌
Long-term persistent memory	✅ SQLite full-text archive (dual-write, never loses data)	✅ Wiki-style MEMORY.md	✅ MEMORY.md + USER.md	⚠️ CLAUDE.md manual	❌	❌
Cross-session memory	⚠️ session-isolated by default, extensible with external memory	✅ built-in cross-session	✅ built-in cross-session	⚠️ session-isolated by default, extensible with external memory	⚠️ session-isolated by default	⚠️ session-isolated by default

Three-tier conversation memory: (1) Context — latest 16 messages loaded directly into LLM context + periodic summary; (2) ToriiDB — self-developed embedded vector database (pardnchiu/ToriiDB) for semantic similarity search on recent conversations; (3) SQLite FTS5 — full-text archive via pardnchiu/go-sqlite, dual-written on every message, never loses data even after history compaction. search_chat_history routes by mode: semantic → ToriiDB, keyword → SQLite FTS5.

---

10. Dependencies & Deployment

	Agenvoy	OpenClaw	Hermes Agent	Claude Code	Codex CLI	Gemini CLI
Direct external dependencies	12	large (pnpm monorepo)	30–40 core + 60+ optional	50+	40+	40+
Self-maintained ecosystem packages	6 (go-bot / go-pkg / go-scheduler / ToriiDB / go-faas / KuraDB)	0	0	0	0	0
Runtime	Go (static binary)	Node.js	Python	Node.js	Node.js + Rust	Node.js
Deployment	single binary	npm install	pip + docker/VPS	npm install	npm install	npm install

---

Where Agenvoy Stands

Dimension	Detail
Clear advantages	Single Go binary, 12 dependencies, self-maintained ecosystem (pardnchiu universe), dispatcher model routing, native platform UI (real buttons/modals), OTP verification, cross-session send to Telegram/Discord from any session, API tool auto-discovery, format reference as lazy-load tool, local-only scheduler (no cloud required)
On par with competitors	Telegram/Discord daemon, TTS/STT, scheduler output push, Skill system, MCP, browser automation, inbound attachment handling, provider coverage (compat layer covers any OpenAI-compatible endpoint)
Where competitors lead	Hermes context compression engine (token-budget compaction: head preservation + middle-turn summarization + iterative recompression, vs Agenvoy's reactive trim-only), OpenClaw 24+ platforms, Hermes MCP server mode, Hermes local STT, OpenClaw/Hermes built-in cross-session memory, Claude Code Computer Use beta, Claude Code cloud cron/task
Codex CLI	Fewest features — CLI + TUI + OpenAI OAuth only, no daemon, no chat platforms, no scheduler

Wiki

• Getting Started
• Architecture
• Core Concepts
• Providers
• Tools
• Memory System
• Skill System
• MCP Integration
• Security and Sandbox
• CLI Reference
• Configuration

License

This project is licensed under the Apache License 2.0.

Community Contributors

Contributor

Just open an issue to share an idea.

Star History

When the curve trends up — that's the signal we want to see. Hit ★ to push it along.

---

©️ 2026 邱敬幃 Pardn Chiu

---