Octop: Self-Hosted AI Assistant with Multi-User Support
Octop earns its keep once you need real multi-user isolation, IM bridges to Chinese workplace chat platforms, or a team of specialized agents under one roof. Setup friction is the catch โ uv provisioning, wizard init, and per-channel credential management demand more than a one-click app. If your team runs on Feishu, DingTalk, or WeCom and you refuse to send data to the cloud, this is the only self-hosted option with first-class IM integration. Otherwise, the overhead outweighs the gains.
Why self-host an AI assistant?
Cloud AI assistants lock your conversations and data to vendor servers, offer no per-user isolation for household or team deployments, and have zero native integration with workplace chat platforms that millions of people use daily. You end up copying outputs manually between ChatGPT and DingTalk. Octop solves this by keeping everything local under `~/.octop/` while routing every surface โ web, CLI, IM, and cron โ through a single control plane with JWT-based multi-user separation.
What is Octop?
Octop is a self-hosted AI assistant platform that runs a single Python process to serve a web dashboard, CLI, IM channels, and cron automation simultaneously, all sharing one control-plane database. Built on the Harness stack, it routes every surface through one in-process processor. You get multi-user JWT isolation, MBTI persona agents, an expert library auto-scanned at boot, RAG over private documents, and native bridges to Feishu, DingTalk, QQ, Discord, and WeCom.
Core features that set Octop apart
- โMulti-user expert team with JWT isolation โ one admin, shared household or team; each user gets their own agents and workspaces
- โMBTI persona agents โ 16 personality templates plus custom system prompts give each agent a distinct character
- โRAG knowledge base โ semantic retrieval grounds agent answers in your private document corpus
- โIM integration โ native bridges to Feishu, DingTalk, QQ, Discord, and WeCom for unified team messaging
- โACP bidirectional integration โ delegate coding tasks to OpenCode or Claude Code with permission gates
- โExpert library auto-scanned at boot โ specialized agent configurations land in `infra/agents/experts/library/` and are immediately available
- โPluggable backends โ local disk, Docker containers, PostgreSQL, or COS/S3 for workspace storage
- โShell command guardrails and PII redaction โ security controls prevent data leakage through tool execution
What can you do with Octop?
- โขFamily or household assistant โ one admin account, each member gets their own agents and memory; specialists can be swapped per task
- โขDeveloper boost โ delegate coding tasks to Claude Code or OpenCode via ACP, troubleshoot from the terminal with AI assistance, or use Browser AI+ for web automation
- โขTeam helper โ multiple agents collaborate in parallel; route tasks from Feishu, DingTalk, or WeCom group chats into coordinated agent workflows
- โขScheduled automation โ configure cron jobs in natural language; daily, the agent executes or triggers jobs on schedule
- โขRemote desktop โ featuring live screen sharing and dashboard input for headless machines on Linux, Windows, and macOS
Getting Octop running locally
Run the one-liner for your platform: `curl -fsSL https://finnie-1258344699.cos.ap-guangzhou.myqcloud.com/octop/install.sh | bash` on macOS/Linux, or `irm https://finnie-1258344699.cos.ap-guangzhou.myqcloud.com/octop/install.ps1 | iex` in PowerShell on Windows. The installer provisions Python 3.12 via uv in an isolated venv under `~/.octop/` and places `octop` on your PATH. Browser automation requires `--extras browser`; Feishu support needs `--extras channels-feishu`. After installation, run `octop init` for the interactive setup wizard, then `octop run` to start.
Strengths
- โOnly self-hosted option with first-class IM bridges to Feishu, DingTalk, and WeCom
- โSingle-process architecture means no external message broker or queue to maintain
- โMulti-user JWT isolation keeps household or team data properly separated
- โExpert library auto-scanned at boot makes specialized agents immediately available
- โACP integration lets agents delegate to Claude Code or OpenCode for real IDE workflows
- โData stays local under `~/.octop/` by default โ no vendor lock-in for storage or LLM providers
Known limitations
- โณInteractive API docs are disabled by default and require editing config.json to enable โ not discoverable for API-first users
- โณExpert library has minimal UI guidance; discovering and configuring experts requires digging into `infra/agents/experts/library/`
- โณPer-channel credential setup (Feishu App ID, Discord bot token, etc.) is non-trivial for non-technical users
- โณThe expert sharing and shared resource pool features are on the roadmap but not yet implemented
- โณBrowser automation (Playwright) and local embedding models require separate install flags โ not included by default
Alternatives to consider
Frequently asked questions
Yes, Octop is MIT licensed and the source is on GitHub. You can run it locally, inspect the code, and self-host it without paying anything. Optional managed services or enterprise support may appear in the future, but the core remains open.
All data stays local under `~/.octop/` by default. Conversations, agent workspaces, credentials, and uploaded documents never leave your machine unless you configure an external storage backend or LLM provider that sends data out. JWT multi-user isolation and PII redaction add additional safeguards for shared deployments.
Yes, via ACP (Agent Client Protocol). You can configure Octop as an ACP server with `octop acp --agent main` for Zed or OpenCode to connect to, or enable the outbound runner per agent to delegate coding tasks to Claude Code or OpenCode from the dashboard. Permission gates control what gets delegated.
Octop works with OpenAI-compatible APIs, DashScope (Qwen models), and Ollama. You configure providers and models per agent in the dashboard or via the CLI. The setup wizard and `octop models` command help with provider configuration.
Each user can run multiple agents, and each agent has its own workspace, providers, channels, and cron schedule. An expert library in `infra/agents/experts/library/` is auto-scanned at boot and made available to switch between specialists per task. The roadmap includes AgentTeams for autonomous multi-agent orchestration.
LangChain is a library for building LLM applications in code; Octop is a deployed platform with a web dashboard, CLI, and IM channels ready to use out of the box. Octop targets households and teams wanting a self-hosted AI assistant, not developers building custom LLM pipelines. It composes the Harness stack (harness-agent, harness-gateway, harness-memory) into a single-process server.
Who should try it โ and who should skip
Octop is worth deploying if you need a self-hosted AI assistant for a household or small team that communicates over Feishu, DingTalk, or WeCom, and you want per-user isolation with a shared control plane. Developers who want to delegate coding tasks to Claude Code or OpenCode from a local agent will benefit from ACP integration. Skip it if you want a simple single-user chatbot with minimal setup, or if your team uses only Western IM platforms and has no requirement for local data hosting.
Related repositories
Curious whether Octop is right for you?
Let ChatGPT, Claude, or Perplexity look into it โ click below and see what AI actually says about Octop.
