Agenvoy Wiki

Agenvoy is a local AI agent that can build, test, and reuse its own tools. A single Go daemon talks to every major LLM provider through one interface, reaches you through Telegram / Discord / TUI / browser, schedules recurring work, searches your local files, and exposes its whole tool library to other agents over MCP.

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.

v1.0.0 is the first formal release and moved the project to AGPL-3.0 with commercial licensing available.

Highlights

• Thirteen provider entries across eleven vendors — OpenAI, Codex, Claude, Gemini, Grok, Grok (xAI), GitHub Copilot, DeepSeek, Mistral, NVIDIA NIM, Ollama Cloud, OpenRouter and Cloudflare, all behind the external go-llm-router module. OpenAI and Grok each carry both an API-key and an OAuth route (Codex, xAI subscription). Beyond those, any OpenAI-compatible /v1 endpoint plugs in through compat (Local/Custom) — Ollama, LM Studio, vLLM, a self-hosted gateway — as many as you want, with models discovered from the endpoint
• Dispatcher-based routing — a dispatcher LLM routes each task to the best-fit worker, with per-model cooldown and fallback when one provider fails. Each model can carry a tier (S / A / B / C, or pass to keep it out of auto routing and subagents)
• Three-pass concurrent tool dispatch — read tools fan out concurrently; write tools stay serial for safety
• Multi-layer memory — rolling summary (incremental, timestamp-cursored) + 24-message recent history + keyword/semantic dual search + cross-session error memory with 90-day TTL, plus an in-run compaction pipeline sized to each model's context window
• Native document RAG — KuraDB registered as an MCP server (mcp__kura__*), installed on its own and added through /mcp
• Skill system — loadable markdown skill packs triggered by /skill-name or run_skill, discovered from Agenvoy's own storage and from Claude / Codex / OpenCode / OpenAI skill directories
• Tool self-extension — the agent writes, sandbox-tests, and version-controls new script_* / api_* tools when nothing existing covers a request
• OS sandbox — Linux bubblewrap / macOS sandbox-exec; argv-only commands, a denied_command denylist, parsed sh -c validation. $HOME is writable with no setup; anything outside it takes one password-backed prompt scoped to that session and path
• MCP client and server — one package on the official modelcontextprotocol/go-sdk; live tool-list refresh in, mcp__<server>__<tool> registration out
• Chat platform integration — Telegram (6-digit OTP first-contact verification) + Discord (native select menus / modals); cross-session push via send_to_chatbot
• Hands-free voice in the browser — the web dashboard listens through a VAD mic loop, wakes on a spoken wake word (hey / 嘿 / 哈囉 ...), speaks replies through the configured text-to-speech model, and lets you cut in mid-sentence to stop it. Spoken turns land in the chat transcript like any other message, so the full written result is still there. This is the browser feature; the TUI's old /voice toggle is gone (v0.34.3)
• Audio & attachments — speech-to-text and text-to-speech are routed independently to an OpenAI or Gemini model; read_files transcribes inbound audio and video, generate_audio writes a .wav. Inbound attachments are saved to the download dir
• Subagents — subagents(mode=invoke) runs in-process under a collection-only charter, at most three legs concurrently, with usage rolled up to the parent session
• Scheduler — cron / one-shot tasks, fsnotify hot-reload, output pushed back to Telegram / Discord
• Built-in web dashboard — the browser UI is embedded in the binary and served by the daemon itself at http://127.0.0.1:17989; sessions, chat, monitors, schedules, models, MCP, rules and notes all run against your own machine
• Local HTTP API — an OpenAI-compatible completion endpoint plus localhost-only management of models, sessions, MCP servers and their OAuth logins, schedules, rules, operator notes, allowlists, and launch-on-login

Source

• Repository: pardnchiu/Agenvoy