Agenvoy - Taiwan-Built Self-Hosted AI Agent Harness & MCP Server

Open Source · AGPL-3.0 · Built in Taiwan

A Personal AI Agent
& MCP Server on Your Machine

A self-hosted agent harness on your own machine — say what you need in one sentence, and it builds the tool, tests it, and reuses it.
Every tool is sandboxed and shared with Claude Code, Codex, and any MCP agent.

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

agen

❯ What's the weather in Taipei?

⏵ Search Tools(weather)

⏵ Plug Tool(Script fetch_weather)

⏺ Currently 28°C in Taipei, partly cloudy with 72% humidity.

Wind from the northeast at 12 km/h.

⏜ 3.2s · claude-sonnet-4-6 · ↑2.1K · ↓340

❯ Report TSMC stock price every morning at 8am on Telegram

⏵ Plug Tool(Script fetch_yahoo_finance)

⏵ Add Schedule(0 8 * * *)

⏺ Done. Scheduled daily at 08:00 — results push to Telegram.

⏜ 5.8s · claude-sonnet-4-6 · ↑3.4K · ↓520

AI Agent Harness

Taiwan-developed AI Agent Harness

Agenvoy is a Taiwan-developed AI Agent Harness built to turn conversation into completed work on your computer. It coordinates models, context, tools, real-time data, task routing, memory, schedules, and execution in one workflow, while keeping control of your files and environment in your hands. Through the Web interface, it also supports hands-free voice interaction with natural speech, wake-word detection, and interruptible spoken replies; the full result remains available in the chat.

Web Dashboard

Drive Your Agent From the Browser

Manage sessions, tools, schedules, and memory from a browser. The dashboard ships inside the binary — start the daemon and open http://127.0.0.1:17989. It is served by your own machine, so nothing leaves your device.

Features

Everything You Need in One Binary

Auto-built tools and a built-in MCP server at the core — plus multi-model routing, scheduling, and memory, all in a single Go binary.

Auto Tool Generation

Ask for anything. If no tool exists, the agent builds one — writes the script, tests it in a sandbox, and registers it. Next time, it's already there.

MCP Server & Client

Server mode shares every sandboxed tool with Claude Code, Codex, and any MCP client — which can auto-build new tools too. Client mode calls external MCP services.

Multi-Model Dispatch

Eleven providers behind one interface, plus any OpenAI-compatible endpoint. The dispatcher routes each task to the best model automatically.

Scheduler

Cron jobs and one-shot tasks with skill binding. Daemon fires them, hot-reloads on change, auto-disables on failure.

3-Tier Memory

Context window, semantic vector search, full-text SQLite archive. Tool failures and their fixes persist across sessions.

Sandboxed Execution

Every command and script runs inside an OS-native sandbox — bubblewrap on Linux, sandbox-exec on macOS.

Skill System

Loadable markdown instruction packs with YAML frontmatter. Slash-command activation, natural language trigger, or cron-bound.

Self-Improvement

When a tool fails, the agent rewrites it, sandbox-tests it, and retries — the fix and whether it worked carry into later sessions. Every file the tools touch is snapshotted, so a version or a whole task can be restored.

Hands-Free Voice

The browser dashboard listens on a VAD mic loop, wakes on a spoken wake word, speaks its reply, and lets you cut in mid-sentence. Every spoken turn still lands in the chat transcript.

RAG Integration

Optional KuraDB, registered as an MCP server for external document search. Keyword + semantic dual-track over your personal knowledge base.

How It Works

Capability Gap → Auto-Build

When no tool fits a request, the agent builds one in real-time — finds an API, writes the script, tests it, and saves it for reuse. No code required.

Step 1

Detect Gap

User asks for live data. No matching tool found.

→

Step 2

Find API

Search public APIs, fetch docs, pick the best match.

→

Step 3

Write & Test

Generate script + schema. Run test_tool to verify.

→

Step 4

Call & Share

Execute the new tool. Auto-persisted for all agents.

MCP

Share Tools Across All Your Agents

One config line gives Claude Code, Codex, and any MCP agent your entire sandboxed tool library — and they can auto-build new tools that everyone instantly shares.

Any MCP-compatible agent connects via stdio and gets access to all tools — built-in, script, API, and extension.

// claude_desktop_config.json
{
  "mcpServers": {
    "agenvoy": {
      "command": "agen"
    }
  }
}

Claude Code Codex OpenCode Cursor

Client

Call External Services

Connect third-party MCP servers via stdio or HTTP/SSE. Tools auto-registered and available to the agent immediately.

// ~/.config/agenvoy/mcp.json
{
  "github": {
    "command": "npx",
    "args": ["@modelcontextprotocol/server-github"]
  }
}

stdio HTTP/SSE Hot Reload

Integrations

Talk to Your Agent Anywhere

Same agent, same tools, same memory — accessible from terminal, chat apps, HTTP, or other AI agents.

TUI

Interactive terminal with slash commands and session switching

OTP-authenticated bot with HTML output and voice input

Discord

Guild-verified bot with markdown and button confirms

REST API

OpenAI-compatible chat completions endpoint

MCP Server

Expose tools to Claude Code, Codex, and more

Providers

11 LLM Providers, One Interface

Eleven model providers behind one Agent.Send() interface, plus as many OpenAI-compatible endpoints as you want — Ollama, LM Studio, vLLM or self-hosted. OAuth routes — Codex, Copilot, xAI subscription — are authentication paths, not extra providers. The dispatcher auto-routes each task to the best model.

Anthropic

Opus · Sonnet · Haiku

OpenAI

GPT-5.x · Codex OAuth

xAI Grok

API key · xAI OAuth

+ OpenAI-compatible

Ollama · LM Studio · vLLM

Tools

32 Built-in Tools, Infinitely Extensible

Related actions sit behind one name and a mode argument instead of a dozen near-identical tools. 25 always register; pkg_manage is added on Linux, find_note once an operator note exists, and four more once their credential or setting exists. Fifteen ship with full schemas — the rest arrive as a name and a description and load their parameters on first use, keeping the initial payload well under the full registry. Missing a tool? The agent builds it.

Get Started

Up and Running in 60 Seconds

One command to install, one command to start. Add a provider and you're ready.

Install

Single binary. macOS and Linux.