Octop is the self-hosted multi-agent assistant that actually plugs into where you already talk
Octop trended on GitHub today, picking up 571 stars in a day on its way past 3,900 with 405 forks. MIT licensed, version 1.0.0, from TencentCloud, at https://github.com/TencentCloud/Octop. The description is a smarter self-hosted AI assistant, multi-user and multi-agent, and the shape of it is a single Python 3.12 FastAPI process that serves a web dashboard, a CLI, several IM channels and background automation all at once.
The design choice that separates it from the pile of agent frameworks is that it does not try to be a place you visit. It bridges into Feishu, DingTalk, Discord, WeCom and QQ, which means the agent lives where the conversation already is rather than asking your team to adopt another tab. On top of that it does web automation, terminal assistance and browser-based remote desktop. The harness stack is split into agent runtime, gateway bridging, memory, and browser automation, and it talks ACP, the Agent Client Protocol.
State lives in SQLite by default, Postgres if you want it, under ~/.octop/, with multi-user JWT auth and PII redaction on the local-first path. That combination is the actual product: multi-user means it is aimed at a team or a household rather than one developer's laptop, and multi-user plus self-hosted plus PII redaction means somebody thought about what happens when six people's messages land in one agent's memory. The RAG knowledge base and plugin ecosystem are table stakes at this point.
The one thing I would raise an eyebrow at is the MBTI personality templates for agents, which is pop psychology wearing a config file. Ignore that and what is left is a coherent answer to a question a lot of teams are asking badly right now: how do you run agents for a group of people without handing a third party your chat history and your terminal. Being under a Tencent Cloud org and MIT licensed at the same time is an unusual combination worth noting, since the license is the part that survives a strategy change.
← Back to all articles
The design choice that separates it from the pile of agent frameworks is that it does not try to be a place you visit. It bridges into Feishu, DingTalk, Discord, WeCom and QQ, which means the agent lives where the conversation already is rather than asking your team to adopt another tab. On top of that it does web automation, terminal assistance and browser-based remote desktop. The harness stack is split into agent runtime, gateway bridging, memory, and browser automation, and it talks ACP, the Agent Client Protocol.
State lives in SQLite by default, Postgres if you want it, under ~/.octop/, with multi-user JWT auth and PII redaction on the local-first path. That combination is the actual product: multi-user means it is aimed at a team or a household rather than one developer's laptop, and multi-user plus self-hosted plus PII redaction means somebody thought about what happens when six people's messages land in one agent's memory. The RAG knowledge base and plugin ecosystem are table stakes at this point.
The one thing I would raise an eyebrow at is the MBTI personality templates for agents, which is pop psychology wearing a config file. Ignore that and what is left is a coherent answer to a question a lot of teams are asking badly right now: how do you run agents for a group of people without handing a third party your chat history and your terminal. Being under a Tencent Cloud org and MIT licensed at the same time is an unusual combination worth noting, since the license is the part that survives a strategy change.
Comments