Tencent's Octop: Open-Source Local AI Workspace with Multi-Agent Architecture

AI assistants can write articles, generate code, and operate browsers, yet they struggle to become long-term personal assistants that retain context across sessions, schedule recurring tasks, or collaborate in multi-user environments.

Octop Overview

Tencent Cloud's Octop (7.9k GitHub stars) addresses these gaps with a fully open-source local AI workspace. Its four core components—Octop Harness, Octop Memory, Octop Browser, and Octop Gateway—are each independently open-sourced, allowing developers to deploy the full stack or adopt individual capabilities.

Technical Architecture

Octop is primarily written in Python. The backend uses FastAPI and uvicorn; the frontend uses React, TypeScript, Vite, and Ant Design. A unified HarnessProcessor handles requests from the web console, instant-messaging platforms, and scheduled tasks. The system runs in a single process without external message queues or brokers. The control-plane database defaults to SQLite with optional PostgreSQL support, reducing initial deployment complexity while preserving extensibility for multi-user, plugin, and external-storage scenarios.

Core Components

Octop Harness : Agent runtime providing model routing, tool calling, skills management, and conversation checkpoints.

Octop Gateway : Normalizes incoming requests from web, IM, and other channels.

Octop Memory : Manages long-term memory storage and retrieval.

Octop Browser : Supplies CDP-based browser automation.

Multi-Agent Expert System

Instead of a single general-purpose assistant, Octop lets users create specialized AI experts—e.g., a writing expert tuned to personal style, a coding expert for development and debugging, a research expert for document management and summarization. Each expert can have its own model, workspace, skills, and message channels. An expert library and sharing mechanism enable reuse and distribution of expert configurations within the same deployment.

RAG Knowledge Base

Built-in retrieval-augmented generation allows agents to answer questions using private documents, technical specs, and internal knowledge, reducing manual lookup time. Knowledge bases can be shared across users and experts in the same deployment.

Tool Integration & Automation

Octop extends agents beyond text generation:

Browser AI+ : Chromium-based automation for web navigation, screenshots, form handling, and browser tasks.

Terminal AI+ : Interactive shell in the web console where AI helps interpret commands, troubleshoot, and execute operations.

Remote Desktop : View and control desktop sessions on Linux, Windows, and macOS from the console.

ACP (Agent Client Protocol) : Bidirectional integration—external IDEs/terminals can consume Octop agents, and Octop can delegate coding tasks to OpenCode, Claude Code, Codex, etc., acting as a unified task entry point.

Multi-User & Multi-Channel Support

JWT-based authentication with admin roles and user isolation enables shared deployments for families or small teams. Users own separate experts, workspaces, and session data while optionally sharing expert configs and knowledge bases. Interfaces include a web dashboard, desktop client, and connectors for Feishu, DingTalk, WeChat, QQ, and Enterprise WeChat, plus HTTP, SSE, and WebSocket APIs. APScheduler powers periodic workflows, turning Octop into a continuously running AI work platform rather than a reactive chat window.

Installation

macOS / Linux:

curl -fsSL https://finnie-1258344699.cos.ap-guangzhou.myqcloud.com/octop/install.sh | bash

Windows (PowerShell):

irm https://finnie-1258344699.cos.ap-guangzhou.myqcloud.com/octop/install.ps1 | iex

Initialize: octop init Run:

# Foreground (API + web console)
octop run

# Custom host/port
octop run --host 0.0.0.0 --port 8088

# Register as system service (systemd / launchd / Windows service)
octop service start

After starting, access http://127.0.0.1:8088 and log in with the admin password set during initialization. Configure LLM providers and API keys to begin using the workspace.