agentcourier

io.github.alhermann由 GitHub 帳號 alhermann 發布v2.2.0更新於 Nov 7, 2026

Let AI agents on different machines message each other and share files, end-to-end encrypted.

已驗證STDIO僅桌面溝通與協作AI 與機器學習

概覽

AI 產生的概覽

讓不同機器上的 AI 代理透過端對端加密的配對連線互相傳訊息與檔案,不需要伺服器或雲端。

可以做什麼
  • 依名稱或位址向另一個代理傳送訊息,並傳檔案給它
  • 讀取收件匣中的來訊與收到的檔案
  • 列出目前可連線或曾經出現過的對端
  • 重新命名目前的代理工作階段,方便他人定址
  • 離線代理的訊息會排隊等它上線;在 Claude Code 中訊息可直接出現在執行中的工作階段
適用情境
當你同時在筆電、HPC 叢集、同事的機器等多處執行編碼代理,希望它們互相提問或傳遞檔案而不必手動複製時適用。
執行需求
每台機器需本機安裝(pip install agentcourier,或以 cargo 安裝)。機器之間需用一次性邀請碼配對一次,並具備網路通路,例如 SSH 通道或自建中繼。未宣告需要帳號、API 金鑰或環境變數。
安裝前請注意
收到的訊息視為不可信文字,代理被要求先徵得使用者同意再行動,但訊息仍可能影響代理行為。收到的檔案會做惡意程式篩檢,但仍需謹慎處理。Claude Code 需加 --dangerously-load-development-channels 參數才能接收推送訊息,該功能為研究預覽。配對即授予對方機器存取權,直到你撤銷。

安裝

在 SourceWeft 中

  1. 開啟 儀表板中的 agentcourier,將其新增到工作區。
  2. 為需要使用其工具的對話啟用該服務。

Desktop only,透過 STDIO。 STDIO 服務會啟動本機處理程序,因此需要 SourceWeft 桌面主機。

其他 MCP 客戶端

參照 儲存庫 中的啟動說明。

README

[agentcourier]

agentcourier

Let your AI agents talk to each other, securely, across machines.
An MCP server for agent-to-agent messages and files between Claude Code, Codex, GitHub Copilot and any other MCP client,
on your laptop, an HPC cluster or a colleague's machine. End-to-end encrypted, no server, no cloud.

[CI] [PyPI] [Post-quantum hybrid encryption] [License]


What it does

  • Messages and files between agents. An agent on your laptop asks the agent on the cluster to check a simulation; the answer comes back into the same conversation. Files can be sent too.
  • Pair once, then it just works. Machines are paired once. Agents can close and restart; they keep their names, and messages for an agent that is offline wait until it is back.
  • Works with running sessions. In Claude Code, messages appear directly in the session. Other clients read them with one tool call, and you get a desktop notification.
  • Only machines you paired. No accounts, no server, no cloud. Every connection is encrypted end to end, including protection against future quantum computers.
  • Guardrails built in. Personal data and secrets are masked, files are screened for malware, and agents are told to treat incoming messages as untrusted.

To an agent, agentcourier is a normal MCP server with the tools courier_send, courier_send_file, courier_inbox, courier_peers and courier_set_name.

Get started

1. Install on each machine (one small program, no admin rights needed):

sh
pip install agentcourier

Or, with Rust: cargo install --git https://github.com/alhermann/agentcourier.

2. Pair two machines once. On the machine the other one connects to (for example your cluster):

sh
agentcourier invite laptop

It prints a one-time code. On the other machine, paste it and say how to reach the first one (here through an SSH tunnel):

sh
ssh -N -L 17312:127.0.0.1:7312 you@cluster-login-nodeagentcourier join agentcourier-… hpc 127.0.0.1:17312

Tip: name machines after their owners (agentcourier invite jana), so you can say "ask Jana's agent".

3. Add agentcourier to your agents:

Claude Code
sh
claude mcp add --scope user agentcourier -- agentcourier mcp --channelclaude --dangerously-load-development-channels server:agentcourier

--channel makes messages appear directly in the running session (Claude Code channels, research preview).

To be woken up by messages, Claude Code must be started with --dangerously-load-development-channels server:agentcourier every time. Channels are a research preview, so Claude Code only accepts them with this flag. Without it, agentcourier still works, but the session only sees new messages when it calls courier_inbox. Tip: add an alias, e.g. alias claude-courier='claude --dangerously-load-development-channels server:agentcourier'.

VS Code (Copilot)

Add to your user mcp.json (Command Palette: MCP: Open User Configuration):

json
{  "servers": {    "agentcourier": { "type": "stdio", "command": "agentcourier", "args": ["mcp"] }  }}

Each window is named after the folder it has open.

VS Code cannot push messages into a chat, so also run agentcourier vscode once and reload the window. A small extension then shows a notice and a status bar counter when a message arrives for that window; one click asks Copilot to read it.

Codex

In ~/.codex/config.toml:

toml
[mcp_servers.agentcourier]command = "agentcourier"args = ["mcp"]

If your system cannot find the agentcourier command, use the full path that pip or cargo printed during installation.

4. Talk. Tell your agent, for example: "Ask my Claude on hpc whether the mesh study has finished, and send it results/summary.csv."

Each session gets a clear name from its tool and project, such as copilot-paper or claude-femsim; its full address is claude-femsim@hpc. You can also tell an agent "call yourself fe-runs". Agents look up who is online, and who was seen before, with courier_peers.

Check what is going on

sh
agentcourier peers      # who is reachable, and agents seen beforeagentcourier status     # connections, paired machines, refused attemptsagentcourier log        # recent activity (no message text)agentcourier untrust X  # cut off a machine immediately

Common questions

How do I let two Claude Code sessions talk to each other?

Add agentcourier to Claude Code (step 3). Every session gets a name, such as claude-femsim, and can message any other session by that name, on the same machine or on a paired one.

Can Claude Code, Codex and GitHub Copilot talk to each other?

Yes. agentcourier is a standard MCP server, so agents from different vendors reach each other the same way, for example a Copilot window on your laptop and a Claude Code session on a server.

How do I connect an agent on my laptop with one on a remote server or HPC cluster?

Run agentcourier on both, pair them once, and connect through an SSH tunnel; no firewall changes or admin rights are needed. See the guide, including a setup for two office PCs that only share an SSH server.

What if the two machines cannot reach each other at all?

Run a relay on a third machine that both can reach, such as a group server or a small virtual machine: agentcourier relay. Both machines connect out to it, so they need no open ports or admin rights. You run it yourself, it lets in only the machines you invite, and it forwards encrypted traffic it cannot read. See the guide.

Can my agent talk to a colleague's agent?

Yes, once your two machines are paired. Each of you names the other's machine, so you can say "ask Jana's agent". Incoming messages are treated as untrusted text, and agents ask their user before acting on them.

Is it secure, and does it need a server or an account?

No server, accounts or cloud are involved: only machines you paired can connect, and every connection is encrypted end to end with post-quantum protection. If you add your own relay, it cannot read messages either. Personal data and secrets are masked, and received files are screened for malware. Details and limits are in Security.

How is this different from A2A or multi-agent frameworks?

Frameworks such as AutoGen, CrewAI or LangGraph run several agents inside one program, and A2A connects agent services over the web. agentcourier connects the coding agents you already use, in their running sessions, across machines and people.

Learn more

  • User guide: names and addresses, connecting across networks and through a relay, files, notifications, commands, configuration, clusters
  • Security: how pairing, encryption, masking and file screening work, and their limits

License

Apache-2.0

來源:README.md,提交 9ccb2c8

工具

0
SourceWeft 尚未在沙箱中啟動過該軟體包,因此其工具未知。

版本歷史

2
  1. v2.2.0最新Nov 7, 2026
  2. v2.1.2Nov 5, 2026

評分

已滿足適用檢查項的 50%

該評分衡量此收錄條目的完整程度,依據註冊中心條目、端點與公開來源計算得出。它不是對伺服器本身的評價,也不表示伺服器安全或目前可用:可連通性是帶日期的檢查結果,評分也不包含安裝量或使用者評價。

  • 近 30 天內可連通25 分

    請保持端點能夠回應 MCP 握手;SourceWeft 每 30 天探測一次。

  • 已宣告認證方式10 分

    請在註冊中心條目中宣告憑證請求標頭,或保持端點能夠回應探測,以便判定其認證方式。

  • 已列出工具15 分

    請讓伺服器在無憑證時回應 tools/list,或發布可在沙箱中無憑證啟動的軟體包。

  • 描述必需15 分
  • 圖示10 分
  • 文件15 分
  • 授權條款10 分