👤 For Humans | 🤖 For AI Agents | 中文 | GitHub
C2C Protocol v2.0

Agent-to-Agent Skill Transfer

Skills move between AI coding agents by a human relaying a short Notes block on any IM platform. Works with Claude Code, Codex, OpenClaw, and anything else that runs shell commands.

AES-256 Human-in-the-loop Any IM Zero deps

🤖 AI agents reading this page: this section is for human operators. Skip to § Install (any agent) below — that's where your machine-readable bootstrap begins.

👤 For Human Users

The scenario

You own multiple AI coding agents — Claude Code on one machine, Codex on another, OpenClaw on a VPS, and so on. You talk to each of them through chat (Telegram, WeChat, terminal, whatever works). One agent has a skill you'd like another agent to learn. C2C lets you transfer the skill without setting up any API, without publishing the skill anywhere, and without logging into any machine yourself.

What you do — 4 steps, no shell commands

  1. Teach both agents how to transfer skills. Say to each agent:

    "Go to https://c2cprotocol.org and install the lobster-distill skill."

    The agent reads this page, follows the Install block, configures itself. Do this once per agent.
  2. Ask the sender to share a skill. Say to the sender agent:

    "Use lobster-distill to share my <skill-name> skill with another agent."

    The agent packs, encrypts, uploads, and replies with one message. It has two clearly separated halves: ═══ FOR ADMIN ═══ (a summary for you) and ═══ FOR TARGET AGENT ═══ (the recipe the receiver will execute).
  3. You relay the message. Long-press → forward in Telegram / WeChat / Signal / Discord, copy-paste into email, or even read it aloud on a call — any text channel works. The two agents do not need to be on the same platform.
  4. The receiver installs it automatically. The receiving agent recognizes the FOR TARGET AGENT half, runs the inline commands, and replies "skill installed." That's it.

Things to watch for

Why use C2C instead of …

What is C2C?

C2C is a protocol for transferring a skill — a folder of code + docs — from one AI coding agent to another, by having a human relay a short "Notes" block between them on any IM platform.

Reference implementation: lobster-distill

The canonical implementation is at github.com/qiaoy01/lobster-distill. Everything needed to install and use it is on this page — no cross-page navigation required.

Install (any agent)

Fetch the two shell scripts directly from the qiaoy01 repo:

mkdir -p skills/lobster-distill
cd skills/lobster-distill
curl -fsSL -O https://raw.githubusercontent.com/qiaoy01/lobster-distill/main/share.sh
curl -fsSL -O https://raw.githubusercontent.com/qiaoy01/lobster-distill/main/receive.sh
curl -fsSL -O https://raw.githubusercontent.com/qiaoy01/lobster-distill/main/SKILL.md
chmod +x share.sh receive.sh

Dependencies: openssl, curl, tar (system built-in, no install step).

Send a skill

From the agent's working directory:

bash skills/lobster-distill/share.sh <skill-dir-or-file> "short description"

share.sh packs the target into tar.gz, encrypts it with AES-256-CBC + PBKDF2, uploads to temporary storage (primary: c2cprotocol.org/share, 10-min expiry; fallback: litterbox.catbox.moe, 24h), and prints a Notes block with two clearly separated sections:

The admin forwards the entire block (long-press → forward, copy-paste, or even read aloud on a call) to the target agent over any text channel: Telegram, WeChat, Discord, Signal, Email, SMS, etc.

Receive a skill

The target agent runs the commands inline from the Notes (they're standard curl/openssl/tar), or uses the helper:

bash skills/lobster-distill/receive.sh <URL> <password> <skill-name> tar

The skill is extracted into skills/<skill-name>/.

Workflow

Sender Agent          Human Admin           Receiver Agent
     │                    │                       │
     │ 1. share.sh        │                       │
     │ 2. Print Notes ──→ │                       │
     │                    │ 3. Forward Notes ───→ │
     │                    │    (any IM / email)   │
     │                    │                       │ 4. receive.sh
     │                    │                       │ 5. Skill installed ✅

Agent-specific install paths

All agents use the same scripts. The skills/ directory convention is what differs:

Claude Code

Codex

OpenClaw

Any other agent

Security

Comparison

FeatureC2C (lobster-distill)Google A2APackage registry
Works across IM platforms✅ Any text channel❌ API only✅ Web only
Human oversight✅ Required❌ Direct AI-to-AI❌ Auto-install
Private / unpublished skills✅ Point-to-point❌ Needs API access❌ Public only
Encryption✅ AES-256Depends❌ Plain
DependenciesNone (openssl, curl, tar)SDK + API serverPackage manager + net
C2C 协议 v2.0

智能体之间的技能传授

技能在 AI 编码智能体之间通过人类转发一段简短的 Notes 传授，任意 IM 平台皆可。适用于 Claude Code、Codex、OpenClaw 以及任何能执行 shell 命令的智能体。

🤖 AI 智能体读者注意：本章节是给人类操作者看的。请跳到 § 安装（任意智能体） — 那里才是你需要的可执行引导。

👤 人类用户指南

场景

你有多个 AI 编码智能体 — 一台机器上的 Claude Code、另一台上的 Codex、VPS 上的 OpenClaw 等等。你通过聊天（Telegram、微信、终端，或其他方式）分别和它们对话。其中一个智能体掌握着你想让另一个也学会的技能。C2C 能让你完成这个传授 — 无需对接任何 API，无需把技能发布到任何公开仓库，你也无需亲自登录任何机器。

你要做的 — 4 步，不碰任何 shell 命令

  1. 让两个智能体学会如何传授技能。 分别对每个智能体说：

    "去 https://c2cprotocol.org 按 Install 小节安装 lobster-distill 技能。"

    智能体会读取本页面，按照 Install 小节配置好自己。每个智能体做一次即可。
  2. 让发送方发起传授。 对发送方智能体说：

    "用 lobster-distill 把我的 <技能名> 技能分享给另一个智能体。"

    智能体会打包、加密、上传，然后给你回一条消息。消息里有两个清晰分隔的部分：═══ 管理员请读 ═══（给你的摘要）和 ═══ 目标智能体请读 ═══（接收方会执行的完整指令）。
  3. 你来中转这条消息。 在 Telegram / 微信 / Signal / Discord 里长按 → 转发，或复制粘贴到邮件、短信，甚至电话念给对方 — 任意能发文字的通道都行。两个智能体不需要在同一平台上。
  4. 接收方自动安装技能。 接收方智能体看到 目标智能体请读 那段，执行其中的命令，回复"已安装"。完成。

注意事项

为什么选 C2C 而不是 …

C2C 是什么？

C2C 是一个协议，用于把一个技能（一个代码+文档的目录）从一个 AI 编码智能体传授到另一个，由人类在任意 IM 平台上转发一段简短的 Notes 完成。

参考实现：lobster-distill

官方实现位于 github.com/qiaoy01/lobster-distill。本页包含安装和使用所需的完整说明 — 无需跨页查看。

安装（任意智能体）

直接从 qiaoy01 仓库获取两个 shell 脚本：

mkdir -p skills/lobster-distill
cd skills/lobster-distill
curl -fsSL -O https://raw.githubusercontent.com/qiaoy01/lobster-distill/main/share.sh
curl -fsSL -O https://raw.githubusercontent.com/qiaoy01/lobster-distill/main/receive.sh
curl -fsSL -O https://raw.githubusercontent.com/qiaoy01/lobster-distill/main/SKILL.md
chmod +x share.sh receive.sh

依赖：openssl、curl、tar（系统自带，无需安装）。

发送技能

在智能体工作目录下运行：

bash skills/lobster-distill/share.sh <技能目录或文件> "简要描述"

share.sh 会把目标打包为 tar.gz，用 AES-256-CBC + PBKDF2 加密，上传到临时存储（优先 c2cprotocol.org/share，10 分钟过期；失败 fallback 到 litterbox.catbox.moe，24 小时），并输出带两个清晰分隔块的 Notes：

管理员将整段 Notes 原样转发（长按→转发、复制粘贴，甚至电话念给对方）给目标智能体，任意文本通道皆可：Telegram、微信、Discord、Signal、邮件、短信等。

接收技能

目标智能体可以直接运行 Notes 中的命令（标准 curl/openssl/tar），或使用辅助脚本：

bash skills/lobster-distill/receive.sh <URL> <密码> <技能名> tar

技能会安装到 skills/<技能名>/。

工作流程

发送方              人类管理员             接收方
  │                    │                    │
  │ 1. share.sh        │                    │
  │ 2. 输出 Notes ───→ │                    │
  │                    │ 3. 转发 Notes ───→ │
  │                    │   （任意 IM/邮件） │
  │                    │                    │ 4. receive.sh
  │                    │                    │ 5. 技能已安装 ✅

各智能体安装路径

所有智能体使用同一套脚本，区别只在 skills/ 目录约定：

Claude Code

Codex

OpenClaw

其他智能体

安全说明

对比

特性C2C (lobster-distill)Google A2A包管理器
跨 IM 平台✅ 任意文本通道❌ 仅 API✅ 仅 Web
人类审核✅ 必须❌ AI 直连❌ 自动安装
私有/未发布技能✅ 点对点❌ 需 API 接入❌ 仅公开
加密✅ AES-256取决于实现❌ 明文
依赖无（openssl、curl、tar）SDK + API 服务包管理器 + 网络