reddit · SaaS · 11-09 04:50 54 分 ← 返回列表

Two months into building multiplayer infrastructure for AI agents. Here's what I've learned so far.

A couple of months ago I went all-in on a side project I'd been obsessing over: Bellman ( bellman.sh ). The problem AI agents are single-player. My Claude session can't talk to yours. My Claude Code session on my laptop can't hand context to my chat session on my phone. And if you want agents from different providers working on the same thing, you're copy-pasting between tabs. What I built Bellman gives agents shared rooms. One agent starts a room and gets a join code. Other agent

五维评分

付费意愿

需求真实度

个人可落地

竞争空间

新鲜度

AI 分析

痛点：AI agent 都是「单机版」：不同工具、设备、模型提供方的会话之间无法共享上下文和文件，用户只能在标签页之间复制粘贴。

目标用户：使用 Claude Code / MCP 的独立开发者与多 agent 工作流的小团队

证据：AI agents are single-player. My Claude session can't talk to yours. My Claude Code session on my laptop can't hand context to my chat session on my phone... you're copy-pasting between tabs. 以及 "Trust is the real problem"

变现：按房间/席位订阅（团队版），企业私有化部署，或围绕 MCP 生态做目录收录/分发型服务收费

MVP：两周内做一个 MCP server：join code 创建共享房间 + 共享 key-value/文件存储 + 简单的 propose/vote 决策日志，先做一个公开「Lobby」房间当活演示，再提交到官方 MCP Registry 和 Claude connector 目录获取分发流量

竞品：Bellman (bellman.sh)、Google A2A / Agent2Agent 协议、LangGraph 多 agent 编排、CrewAI、AutoGen

摘要：作者两个月全职做出 Bellman：一个基于 MCP 的「多 agent 共享房间」基础设施，提供 join code、共享画布、共享文件系统和 agent 投票共识机制。他总结分销胜过功能——上架官方 MCP Registry 和 Claude connector 目录带来的曝光超过任何功能迭代；同时指出「信任」而非技术才是真正难点。全帖没有任何付费用户或收入数据，付费意愿未被验证。