Agenvoy

A local AI agent harness in Go that runs 24/7 with auto-built tools, MCP, and Ollama

Automatically checked · GitHub checked 2026-11-13

Agenvoy

Overview

Agenvoy

Your private, local AI agent — running 24/7 on your own machine

Not a VM in someone else's cloud — Agenvoy runs as a daemon on your computer and keeps memory, schedules, tools and credentials on your own disk.
Say what you need in one sentence; it builds the tool, sandbox-tests it, and shares it with Claude Code, Codex and any MCP agent — all from a single Go binary.

agenvoy%2FAgenvoy | Trendshift

Version License

English · 繁體中文 · Docs

[!NOTE] Agenvoy is currently in a review-and-cleanup period. File-by-file progress: Optimization Progress.

Quick start

  1. Install (macOS / Linux; on Windows, run it inside WSL — see below):

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

    The installer opens the TUI and starts the daemon when it finishes; run agen to open it again later.

  2. Add a model — type /model add and sign in with a subscription you already have (Copilot, ChatGPT, Claude) or paste an API key.

Then just ask. New here? From Copilot sign-in to first answer in 5 minutes.

Windows (via WSL)

Open PowerShell as an administrator, then list and install a Linux distribution:

wsl --online --list
wsl --install <distribution-name>

After installation, restart your computer, open a WSL terminal, and run the install command above.

macOS tip: If you run schedules on a MacBook, also run sudo pmset -c sleep 0 so sleep does not interrupt them.

What a session looks like

❯ What's the weather in Taipei?

⏺ Checking for an existing weather tool, then fetching current conditions.

  Ran 2 tool calls

⏺ Currently 28°C in Taipei, partly cloudy with 72% humidity.
  Wind from the northeast at 12 km/h.
  ⎿ 3.2s(41.5t/s)  ↑ 21.4K(11%) ↓ 133  08:12

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

⏺ Loading the finance tool and creating a daily 08:00 schedule.

  Ran 3 tool calls

⏺ Done. Scheduled daily at 08:00 — results push to Telegram.
  ⎿ 5.8s(36.2t/s)  ↑ 34.0K(17%) ↓ 210  08:13

The first request finds an existing script tool and answers with live data; the second binds a tool to a cron schedule whose result is pushed to Telegram.

Taiwan-developed AI Agent Harness

Agenvoy is a Taiwan-developed AI Agent Harness by Pardn Chiu, 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.

Why Agenvoy

A chat can give you an answer; work needs a result. Agenvoy breaks requests into steps, calls tools, and delivers outcomes on your computer—while you retain control of files, tools, schedules, and working context.

  • Turns conversation into deliverable work — Research live data, organize files, and complete multi-step tasks with an agent that acts and reports the result.
  • Fills capability gaps itself — Finds a keyless public API and calls it directly; when none fits, creates, tests, and keeps a new tool, ready to reuse next time.
  • Shares one tool library across agents — Agenvoy, Claude Code, Codex, and other agents use the same sandboxed tools instead of rebuilding them.
  • Keeps automation running — Create schedules in one sentence; recurring work runs in your environment and pushes the result.
  • Makes every step visible and controllable — Command output streams to the TUI and Web dashboard, while sensitive paths and restricted actions still require confirmation.
  • Connects models and external services freely — Route models by task and configure image generation, STT, TTS, stdio/HTTP MCP servers, and OAuth.
  • Provides private access from anywhere — The local daemon connects outward to Telegram and Discord, without making your host public or opening inbound ports.

Keep your keys and data on your machine

Meta Muse, OpenAI dots and Grok Bot each run your agent on a cloud VM the vendor operates, with your logins inside it. Agenvoy runs on the computer you already own, so you decide what leaves it.

GuaranteeDetails
Data on your diskSessions, memory and schedules live in SQLite under ~/.config/agenvoy. No Agenvoy account, no hosted copy
Credentials in your keychainAPI keys and OAuth tokens sit in the OS keychain and are used only by the local daemon; secrets in tool output are masked before the model or logs see them
Sandboxed by defaultCommands run under sandbox-exec on macOS or bwrap on Linux with network off unless you approve it for that call; management endpoints answer localhost only
Your choice of modelUse the provider you trust, or point Agenvoy at Ollama or LM Studio so prompts never leave your machine

What you can do with it

Ask live questions and get live answers (Web Search / Tool Generate)

What's the weather in Taipei?

The agent finds current data, calls tools, and gives you the answer.

If a tool doesn't exist, it calls a public API or builds one.

Web Search / Tool Generate demo

Turn one sentence into automation (Scheduler)

Report TSMC stock price every morning at 8am

The agent asks:

  • Where to push results
  • What format you want
  • When to run

Then creates the schedule automatically.

Scheduler demo

Ask questions about your local files (File Search / RAG)

Find all invoices from last year

Which document mentions Prompt guide?

The agent searches your local files and answers directly.

File Search / RAG demo

Finish multi-step work (Skills / Sub-agents)

Summarize today's GitHub commits and generate a progress report

The agent breaks down the task, calls tools, combines results, and replies.

Skills / Sub-agents demo

No plugin to wait for — it builds the tool

When no tool fits a live-data request, the agent first calls a keyless public API directly; when none fits, or your own API needs a key, it writes a script tool, tests it in the sandbox, and saves it for every agent to reuse. No code required.

flowchart LR
  A[Detect gap] --> B[Find API]
  B -->|fits| D[Call and share]
  B -->|none fits or needs a key| C[Write and test]
  C --> D

Give other agents the same ability (MCP Server)

Agenvoy is also an MCP server, so this ability is not only its own. Claude Code, Codex, OpenCode, and other AI agents can connect and:

  • Use all your sandboxed tools
  • Auto-build new tools when none exist
  • Share every tool across all agents

One line of config. Instant shared tool library. Tools created in the demo: fetch_weather · fetch_crypto_price

1. Claude Code creates a weather tool

Claude Code creates a weather tool

2. Codex reuses it and creates a crypto tool

Codex creates a crypto tool

3. Agenvoy tests both tools

Agenvoy tests both tools

Everything in one Go binary

OpenClaw, Hermes Agent, DeepSeek Harness and pi are TypeScript or Python projects; Agenvoy's core is a single Go binary, with auto-built tools and a built-in MCP server at the center.

CapabilityWhat it does
Auto tool generationCalls a public API, or writes, sandbox-tests and registers a tool when none fits
MCP server & clientShares every sandboxed tool with Claude Code, Codex and any MCP client; calls external MCP servers over stdio or HTTP with live tool refresh and OAuth
Multi-model dispatchEleven providers behind one interface plus any OpenAI-compatible endpoint; each task is routed to the model best suited to it
SchedulerOne-shot and cron tasks bound to skills, created from a single sentence
3-tier memoryContext window, semantic vector search and full-text SQLite archive; tool failures and their fixes persist across sessions
Sandboxed executionEvery command and script runs in an OS-native sandbox — bubblewrap on Linux, sandbox-exec on macOS
Skills & sub-agentsLoadable markdown instruction packs and multi-agent collaboration for multi-step work
Self-improvementRewrites and retries a failing tool; every file the tools touch is versioned, so a version or a whole task can be restored
Realtime voiceThe Web dashboard talks with you live through Gemini Live and lets you cut in mid-sentence
Live output & rich chatCommand progress streams to the TUI and Web dashboard; Web chat renders charts, Mermaid diagrams, math and highlighted code
Secret maskingKeys and tokens in tool output are masked before the model or logs see them
Transcription & imagesAudio and video to text; image generation through a configured provider

37 built-in tools: related actions sit behind one name and a mode argument, and only seven schemas ride along with every run — the rest are fetched on demand, so the prompt cache survives the whole session. See Built-in Tools.

Talk to your agent anywhere

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

ChannelHow you use it
TUIInteractive terminal with slash commands and session switching
Web dashboardShips inside the binary at http://127.0.0.1:17989; manage sessions, tools, schedules and memory
Telegram / DiscordThe local daemon connects outward with a bot token — no inbound port, no reverse proxy. New messages steer a running task; /new <message> starts a separate one
REST APIOpenAI-compatible chat completions endpoint, plus agen -p for one-shot prompts
MCP serverExpose tools to Claude Code, Codex, OpenCode and Cursor

Agenvoy Web Dashboard demo

See the Web UI comparison (old / new, light / dark) for the v1.3.4 interface refresh.

Providers

Eleven model providers behind one interface, plus as many OpenAI-compatible endpoints as you want. Subscription routes — Claude Code, Codex, Copilot, xAI — sign in with the plan you already pay for instead of an API key.

ProviderModels / route
AnthropicOpus · Sonnet · Claude Code
OpenAIGPT-5.x · Codex OAuth
GeminiPro · Flash
GitHub CopilotGitHub OAuth
NVIDIA NIMNIM API
xAI GrokAPI key · xAI OAuth · searches X directly
DeepSeekFlash · Pro
MistralAPI key
Ollama CloudAPI key
OpenRouter200+ models
CloudflareWorkers AI
+ OpenAI-compatibleOllama · LM Studio · vLLM or self-hosted

Run one prompt from the shell

agen -p "Summarize today's commits" --stream

agen -p sends one prompt to the local daemon and prints the reply. It is read-only and sandboxed by default; add --allow-edit to let it write and --skip-sandbox to let it leave $HOME. --skip-sandbox (enable_sandbox:false on /v1/send) is very dangerous: it allows sudo and paths outside $HOME with no confirmation, and the daemon accepts it only from this machine. See the docs for every flag.

Recommended setup

[!NOTE] No free plan is recommended. Since 2026-11-09 (5:40 PM PT), Ollama Cloud's Free plan is pay-as-you-go: hosted models such as gemma4:31b, gpt-oss:120b, gpt-oss:20b, nemotron-3-super and nemotron-3-ultra need purchased credits, so Ollama Cloud no longer offers a free way to try Agenvoy. Running models locally with Ollama stays free, but small local models call tools unreliably. The lowest-cost recommendation is GitHub Copilot ($10/mo) below.

NVIDIA NIM's free API is not recommended either: NIM has not offered gpt-oss-120b since 2026-09-04, and since then most of the remaining models time out or respond slowly, and none meets the minimum response speed and basic reasoning an agent needs.

A cost-effective model setup to get started:

  1. Choose a subscription model for everyday primary use, such as:

    • GitHub Copilot ($10/mo) — pick the luna model: its quota drains slowly enough for daily use and it is capable enough for most work
    • OpenAI ChatGPT Plus ($20/mo)
    • SuperGrok ($30/mo)
  2. Strongly recommended: Add TypeSafe's Jev as the decision model: create a key in the TypeSafe Console, add it with /model add → API Key → TypeSafe, then pick typesafe@jev-latest under /model → decision. It is billed only on input tokens ($42 per billion, output free, per TypeSafe's model page on 2026-11-11) and keeps irrelevant skills, memory entries, search results, page boilerplate, and unrelated file and history reads out of each turn; on the author's machine it left 95.8% of skill entries out of the turn context. Setup and measurements: Quickstart.

Who it's for

If you want AI to complete work—not merely respond—within an environment you control, Agenvoy is for you:

  • People who want to turn research, file work, and recurring reporting into reusable automation
  • Developers who want a self-hosted agent with local data control and sandbox guardrails
  • Teams that want Claude Code, Codex, and other agents to share tools instead of rebuilding them
  • Technical operators who need private access to a local agent through the Web, Telegram, or Discord

Validated designs

Agenvoy has already implemented the design directions below, and other teams have since adopted similar approaches, confirming that these directions solve real problems agents face in practice. Every milestone links to a public commit or release, so you can check it yourself.

Show the timeline
Shipped in AgenvoyDesignSame direction elsewhere
2025-06-28 Implement core application
2025-06-29 add traditional memory structure
Summary-based memory for unbounded context2025-11-24 Claude client-side compaction
2026-01-04 bubblewrap sandbox
2026-02-01 harden bubblewrap sandbox
2026-03-07 block sensitive paths and credentials
2026-03-18 sandbox execution with bubblewrap
Sandbox mode for secure agents2026-03-16 NVIDIA NemoClaw
2026-02-07 Copilot device code login
2026-04-04 OpenAI Codex (OAuth) provider
Sign in with an existing AI subscription2026-09-29 Pi 0.99.0 Sign in with ChatGPT
2026-02-27 LLM-driven agent routing
2026-03-10 configurable dispatcher model selection
2026-04-15 strengthen agent tier routing
2026-06-10 tiered routing across providers
2026-09-11 configurable model tiers for agent and subagent routing
2026-09-11 model tier controls on fallback priority
Route each task to the model best suited to it2026-09-14 Copilot auto model selection tiers: efficiency / balance / intelligence
2026-09-29 Pi 0.99.0 virtual models: per-request model and thinking level
2026-09-29 Pi 0.99.0 classifier models
2026-03-23 script tool support with sandboxed execution
2026-05-14 script tool scaffolding skill
2026-06-05 script tool runtime metadata
Custom JS / Python tools run in the sandbox2026-09-29 Pi 0.99.0 codemode: model-written JavaScript in a QuickJS sandbox
2026-04-02 deferred tool loading with search_tools
2026-05-05 MCP client adapter with stdio/HTTP transports
2026-06-07 tool search registry
2026-08-08 MCP OAuth flow
Load tools on demand instead of declaring them all2026-09-29 Pi 0.99.0 built-in MCP + tool_search
2026-04-17 invoke_subagent tool
2026-04-25 session action logs
2026-04-25 session log streaming endpoint
2026-04-28 subagent name dispatch
Invoke sub-agents and track their logs2026-08-03 Claude Code cross-session messaging
2026-05-28 extension tool loading
2026-05-28 extension install and upload skills
2026-05-29 extension marketplace
2026-06-08 ext_ extension tools and scaffolding
Everything is a plugin2026-08-13 DeepSeek Harness: everything is a plugin
2026-06-03 ask_user interruption async resume
2026-06-04 preserve in-progress action memory across interruptions
2026-06-29 pending task resume
2026-07-03 write_todo checklist flow
Resume long tasks from a checkpoint2026-08-04 pi harness v2 in-memory storage
2026-08-05 pi indexed harness recovery queries
2026-08-05 pi validate harness recovery record logs
2026-08-05 pi harness v2 jsonl backend
2026-08-06 pi atomic writes + torn-tail truncation
2026-08-14 pi AgentHarness R3 generation recovery
2026-08-14 pi AgentHarness R4 tool execution

Docs

Full documentation at agenvoy.com

License

This project is dual-licensed.

  • Open source — GNU Affero General Public License v3.0. You may use, modify and distribute it, provided derivative works and any network-accessible service built on it are released under the same license, source included.
  • Commercial — for use that cannot meet the AGPL-3.0 source-disclosure requirement, a commercial license is available. It covers rights to the software only; any work on the project requested from the developer is quoted and charged separately. See COMMERCIAL.md.

Under section 7(b) of the AGPL-3.0, the License and Source entries in the Other group of the web interface's Config > System tab must be preserved: they carry the copyright notice, the warranty disclaimer, the license statement, and the corresponding-source link that section 13 requires. They may be relocated or restyled while remaining prominently visible, but not removed or hidden. A commercial license removes this requirement.

Author

Just open an issue to share an idea.

Agenvoy contributors

©️ 2026 邱敬幃 Pardn Chiu