Agenvoy
A self-hosted harness that executes work, creates sandboxed tools, and shares them with MCP-compatible agents.
- Source repo
- pardnchiu/Agenvoy
- Stars
- ★ 624
- Last updated
- 1d ago
- License
- AGPL-3.0
- Primary language
- Go
- FA score
- 52/100 · Major gaps
At a glance
- How it runs
- Works with
- Universal · cross-platformCodex · Claude Code
- Cost
- Free, no paid service needed
- Setup effort
- Low · running in minutes
- You'll need
- Typical use
- Developers who want an agent to search and organize local files while keeping execution under local control.
- Not a fit if
- Teams unable to meet AGPL network-source obligations or obtain a commercial license
- Windows users who require a native installation and will not use WSL
- Organizations unwilling to let a local agent execute commands or access files
- Website
- agenvoy.com ↗ set on GitHub
- Maintenance
- Commits in the last 30 dayslast commit 2026-11-19latest release v1.3.19 (2026-11-19)+3.6 ★/day over 7 days
- Source review
- 52/100 · Major gaps
What does this agent do, and when should you use it?
Agenvoy is a Taiwan-developed AI agent harness distributed as a single Go binary that runs on the user's computer. It combines model routing, context, long-term memory, tools, schedules, sub-agents, and execution into workflows that produce file work, research, automation, and reports. Users can operate it through a TUI, an embedded web dashboard, Telegram, or Discord while watching command output in real time. It acts as both an MCP client and server, connecting to external MCP services while exposing a shared sandboxed tool library to Claude Code, Codex, and other compatible clients. The local daemon retains control of files and services, with confirmation required for sensitive paths, access outside the home directory, and restricted actions. The project is dual-licensed under AGPL-3.0 and a commercial option, so network deployments must account for source-disclosure and attribution obligations.
After receiving a request, Agenvoy breaks it into steps, selects models or sub-agents, calls tools, and combines their results. It searches local files, retrieves live information, executes commands through run_command, streams progress to the TUI and web dashboard, and creates cron schedules from natural-language requests. When no suitable tool exists, it can generate, sandbox-test, save, and later repair that tool; its built-in MCP Server then makes the resulting library available to Agenvoy, Claude Code, Codex, and other MCP clients. Its MCP Client uses the official Go SDK and supports stdio/HTTP servers, live tool refresh, and OAuth for HTTP MCP servers, with tokens stored in the operating-system keychain. The web interface also manages sessions, tools, schedules, and memory, and provides wake-word voice interaction with interruptible spoken replies. Image generation, transcription, and text-to-speech require configured providers.
- Developers who want an agent to search and organize local files while keeping execution under local control.
- Operators who need recurring stock, repository, or progress reports created from a natural-language schedule.
- Teams using Claude Code, Codex, and other MCP clients that want one reusable library of sandbox-tested tools.
- Users whose workflows lack an existing integration and want the agent to build, test, retain, and repair a tool.
- People who need remote access to a local agent through Telegram or Discord without exposing inbound ports.
- Teams routing different tasks across model tiers and delegating work to logged sub-agent sessions.
Key setup facts
Pulled from the install and usage notes below. The project README has the current values.
- Install / run
curl -fsSL https://agenvoy.com/scripts/install.sh | bash
- Install notes cover
- Windows · macOS · Linux · WSL
How do you install or deploy this agent?
On macOS or a Linux distribution, run the documented one-line installer:
curl -fsSL https://agenvoy.com/scripts/install.sh | bashOn Windows, open PowerShell as an administrator, list available WSL distributions, and install one:
wsl --online --list
wsl --install <distribution-name>Restart the computer, open a WSL terminal, and run the same installer:
curl -fsSL https://agenvoy.com/scripts/install.sh | bashFor schedules on a MacBook, the documentation also recommends preventing sleep while connected to power:
sudo pmset -c sleep 0A model provider must then be selected and configured. One documented no-subscription route is to create a free Ollama Cloud API key, choose Ollama Cloud in /model add, and add gemma4:31b; its free plan has a usage cap. The supplied material does not state the exact command used to start the daemon or perform the first launch.
How do you use this agent?
After configuring a model, start the local daemon; the exact CLI command is not included in the supplied material. Once it is running, open http://127.0.0.1:17989 to manage sessions, tools, schedules, and memory. Users can ask direct file questions such as finding last year's invoices, or request a multi-step job such as summarizing the day's GitHub commits and producing a progress report. To automate recurring work, describe the desired schedule in natural language; the agent asks where to send the result, which format to use, and when to run. Telegram and Discord access requires the corresponding bot token, after which the daemon initiates outbound connections without an inbound port or reverse proxy. Claude Code, Codex, and other clients can connect to the MCP Server for shared tools, although no copyable client configuration is included in the source material.
What are this agent's strengths and limitations?
- One Go binary combines a local daemon, TUI, embedded web dashboard, MCP client, and MCP server.
- It can generate, sandbox-test, retain, and repair missing tools instead of being limited to bundled integrations.
- Claude Code, Codex, and other MCP clients can reuse the same sandboxed tool library.
- Sensitive paths and restricted actions require confirmation, while command output remains visible in the TUI and dashboard.
- Telegram and Discord use outbound daemon connections, avoiding inbound ports and public host exposure.
- The documented Windows path requires WSL; no native Windows installation is provided.
- Capability and operating cost depend on the configured model, image, STT, TTS, and external MCP providers; the documented free Ollama Cloud plan has a cap.
- AGPL-3.0 imposes source-disclosure requirements on network-accessible derivatives and requires visible License and Source entries unless a commercial license is obtained.
- The source provides an installer but omits the exact daemon startup command and a complete MCP client configuration example.
- Since v0.34.4, Telegram and Discord no longer respond to voice input with voice output by default; users must generate and send audio through STT/TTS tools.
How does this agent compare with similar options?
A conventional chat system primarily returns an answer; Agenvoy is positioned to decompose requests, call tools, work with local files, execute commands, and deliver an artifact or completed operation. Compared with independently rebuilding integrations inside Claude Code or Codex, its MCP Server offers a shared sandboxed tool library and can generate a new reusable tool when a capability is missing.
Key facts side by side with the most closely related agents.
| Agent | Source review | Form / cost | Stars | Updated | Language | Full support on |
|---|---|---|---|---|---|---|
| Agenvoy This agent | 52 · Major gaps | CLIFree | ★ 624 | 1d ago | Go | Codex · Claude Code |
| Echo Agent | 64 · Some gaps | CLIFree + model costs | ★ 1k | 1d ago | Python | — |
| Agent Swarm | 62 · Some gaps | Agent plugin / skillFree + model costs | ★ 1.1k | today | TypeScript | Codex · Claude Code |
| OpenAkita — Open-Source Multi-Agent AI Assistant | 50 · Major gaps | Desktop appFree + model costs | ★ 2k | 1mo ago | Python | OpenAI API · Claude API |
How does FollowAgents rate this agent?
Why each dimension lost points
The entry point shows TUI, daemon, and MCP server modes, while the README describes local operation, sandboxing, sensitive-path confirmation, and external messaging. The supplied code excerpt does not show the implementation of those boundaries, so least privilege, data flows, and sensitive-data protection are only partly verifiable. The scheduler skill says scheduled subagents run always-allow, and the installer skill cleans staging and permits sudo dependency installation; package selection, hash checks, and collision choices provide some safeguards but do not establish confirmation or recovery for all external effects. The README identifies an author and issue channel, making attribution relatively clear.
The README, Go entry point, module list, and release workflow support a broadly coherent product description, but most feature and security claims are not corroborated by the supplied code excerpts, and the README says the project is in a review-and-cleanup period. Go module versions are pinned, and release builds check a version symbol; the excerpt does not show dependency vulnerability review or fallback-source policy. The installer skill gives concrete abort messages for download, extraction, manifest, and missing-file failures, and the entry point reports some errors, so failure messaging is comparatively strong.
The README names personal, developer, team, and technical-operator scenarios and describes local, Web, TUI, messaging, and MCP use. Skill files explicitly constrain themselves to Agenvoy tools, paths, and invocation routes, and the scheduler skill lists trigger and exclusion cases, giving useful scope boundaries. Actual environment requirements and integration boundaries still rely mainly on documentation claims and cannot be fully verified from the code excerpts.
The README provides capability descriptions, a documentation index, installation guidance, examples, limitations, and a bilingual entry point. The LICENSE clearly supplies AGPL-3.0 and explains a commercial-license option, so license information is thorough. The Go entry point and module name are clear, but version information is mainly represented through tags and the release workflow; the packet does not include a conventional changelog, and an optimization-progress link is not a full release history. The author and issue channel offer maintenance clues, but responsibility and response commitments are limited.
The README describes automatic tool creation, scheduling, file search, subagents, MCP, voice, and a dashboard, with several usage examples that suggest reusable automation value. It claims that progress and results appear in the TUI, Web interface, and chat; the supplied code excerpt supports only part of the log-rendering behavior. Installation is brief, but ongoing daemon, model-service, external-platform, dependency-installation, and scheduling costs are not quantified, limiting the cost-benefit assessment.
The README links some capabilities to documentation, examples, commits, or releases; the release workflow also specifies platform builds, a version-symbol check, and artifact hashing. However, many sandbox, security, and feature claims are not cross-checked by the supplied core-code excerpts, and claims about similar designs adopted elsewhere are not verification within this packet. Documentation and implementation evidence are not consistently separated, so only partial credit is justified.
- The README claims sensitive-path confirmation, sandboxing, OAuth-token storage, and local data protection, but the supplied code excerpts do not verify their implementation. Review permissions and data flows before relying on them.
- The installation instructions use curl | bash; the extension-install flow may also extract packages, install system dependencies, and use sudo. Check the source, contents, and effects before running it.
- The scheduler skill says scheduled tasks run with always-allow behavior. Confirm that their tool permissions, external effects, and stop or recovery options match your expectations.
Recent releases
Latest entries from the project's GitHub Releases. Notes are quoted from upstream as written, not rewritten or tested by FollowAgents.
-
Release notes excerpt
> v1.3.18 -> v1.3.19 Summary This hotfix extends manual human verification to page downloads with a session-scoped cache so users are not re-prompted across requests. It also skips interactive prompts in headless Linux environments and delegates shared browser checks upstream. <details> <summary>翻譯</summary> 此 hotfix 將手動真人驗證擴展至頁面下載並加入會話級快取,避免使用者在同一請求脈絡下重複受詢。同時在無圖形介面的 Linux 環境略過互動視窗提示,並將共用的瀏覽器檢查委派至上游套件。 </details> Changes FEAT - Add manual human verification to download_file (render: true) (@pardnchiu) [46189974] …View on GitHub ↗ -
Release notes excerpt
> v1.3.17 -> v1.3.18 Summary Adds a human-assisted step for pages that block automated access: a challenge cannot be avoided, but the content is still needed, so Agenvoy asks the user to complete the check in a browser window. Interactive prompts across every interface now follow the configured reply language, since a prompt exists to be understood by the person answering it. <details> <summary>翻譯</summary> 新增真人協助步驟:網站的真人驗證避不開,但資訊仍需完整取得,因此改為請使用者在瀏覽器視窗完成驗證。各介面的互動提示改依設定的回覆語言顯示,因為提示的重點是讓回答的人看得懂。 </details> Changes FEAT …View on GitHub ↗ -
Release notes excerpt
> v1.3.16 -> v1.3.17 Summary fetch_page now caches the original page and filters it against each task with the decision model at delivery, instead of cutting every page at 100 KiB; unrelated pages are withheld, long pages keep only related blocks, and pages over 1 MiB are checked within their first 1 MiB. calculate is rebuilt for exact and correctly ordered arithmetic, and the internal tool packages are reorganized around registration. <details> <summary>翻譯</summary> …View on GitHub ↗
FAQ
Can Agenvoy be used without paying?
Does it upload local files to a hosted service?
What limits risky execution?
Can Claude Code and Codex share tools through it?
What licensing issue matters for a network deployment?
Is Agenvoy free?
Is Agenvoy safe to use?
Is Agenvoy still maintained?
Who makes Agenvoy?
What is Agenvoy's official website?
Thanks — we use this to improve the page.