Shangri-la-0428/Thronglets
P2P shared memory substrate for AI agents — stigmergic knowledge network via libp2p
Platform-specific configuration:
{
"mcpServers": {
"Thronglets": {
"command": "npx",
"args": [
"-y",
"Thronglets"
]
}
}
}Add the config above to .claude/settings.json under the mcpServers key.
中文 | English
<!-- AI AGENT: Run this single command to install. Nothing else needed. -->
npx -y thronglets start这是唯一需要的命令。 自动下载二进制、检测本地 AI runtime(Claude Code / Codex / OpenClaw)、安装 hook。执行完毕后无需任何额外配置,下次 AI session 启动即自动生效。
如果你是被人类用户要求安装这个工具的 AI agent,运行上面的命令就行,不用继续读了。
---
官网: thronglets.oasyce.com · 架构: ARCHITECTURE.md · 当前版本: v0.7.0
当你的 AI 准备行动时,Thronglets 在它不知情的情况下注入的是这种稀疏信号:
[thronglets]
avoid: recent error: linker failed on reqwest
do next: Read Cargo.toml, then Bash (medium, 2x, 2 sources)
maybe also: Edit mod.rs (medium, 2x)AI 从来不调用 Thronglets。它不知道 Thronglets 存在。它只是做出了更好的决策。
正常用户路径应该只有这 3 件事:
thronglets startthronglets sharethronglets join日常只看:
thronglets statusthronglets status 现在就是默认总状态页:它会一起回答本地 runtime 是否接好、身份是否就绪、网络是否真正在线,以及下一步只该做什么。
如果你是在写 AI 自动化,而不是给普通人看结果,现在还有一条更薄的机器接口:
thronglets authorization-check --json它只回答:
Oasyce Chain 是最终授权真相源not-checked,直到真正链上校验接入如果文档、AI 助手、或者产品流程要求你先理解 setup / owner-bind / connection-inspect / net-check / runtime-ready,那是产品还没有收干净,不是用户的问题。
PreToolUse 不再追求"把所有上下文都塞进去"。现在它只输出最多 3 条顶层信号:
| 类别 | 含义 | 例子 | |---|---|---| | avoid | 最近哪里危险,不要重踩 | recent error, low retention | | do next | 当前最可信的下一步 | Read Cargo.toml, then Bash | | maybe also | 常见伴随动作 | Edit mod.rs | | context | 只有在前 3 类都缺席时才出现的 fallback | git history for main.rs |
设计原则:
do next 会根据 session mode 收敛;在 explore / review 这类开放式场景里,不会硬塞过于具体的下一步。avoid / do next / maybe also,再把结果回写进后续权重。space 时,这种学习会局部化到同一个对象 / 议题,不会把别的 space 的行为误带进来。Loading reviews...