Self-Hosted AI Assistants Compared: OpenClaw, nanobot, CowAgent and Letta

What a self-hosted personal AI assistant is, what to weigh before running one, and how OpenClaw, nanobot, CowAgent, Letta and the desktop-first alternatives differ.

AgentIndex · Published 2026-10-08 · Updated 2026-10-08

A growing class of open-source projects turns an AI model into an assistant that lives on your own hardware and talks to you through the messaging apps you already use. It remembers things, runs tools, works on schedules, and does not hand your conversations to a hosted service. They look similar from a distance, but they differ in architecture, in how much they assume about your setup and in how much power they hand to the model.

This guide explains what to weigh before running one and compares four of the best-known options in the AgentIndex directory, all of which we have read the current documentation for: OpenClaw, nanobot, CowAgent and Letta. Details are from each project's README as of October 2026.

What "self-hosted assistant" means

The common pattern has four parts:

  1. A gateway or runtime that runs on your computer or server and holds the assistant's state, memory and credentials.
  2. Channels: connections to chat apps (WhatsApp, Telegram, Slack, WeChat, Feishu and others) and sometimes a web UI or native apps.
  3. A model layer that you choose: a hosted provider through your own API key, or a local model.
  4. Tools and skills that let the assistant act: reading files, running commands, browsing, calling APIs.

Compared with a chat website, you gain control over data and cost, and the ability to reach the same assistant from many places. You take on setup, updates and security.

Five things to weigh first

Where do you want to talk to it? If it must live in WeChat or Feishu, check channel support before anything else. Coverage differs a lot between projects.

How much power will you give it? Assistants that can run shell commands and edit files are useful and dangerous in equal measure. Look for sandboxing options, approval modes and clear documentation of what runs on the host.

How will memory work? Some assistants keep conversation history; some maintain long-term memory that they rewrite as they learn; some track memory in files you can read. Decide whether you want inspectable memory or opaque convenience.

How heavy a system do you want? A small, readable Python core you can modify is a different choice from a large platform with native apps and a plugin marketplace.

Who maintains it, and how is it licensed? Check recent activity and read the licence. For a service you rely on daily, an actively maintained project matters.

The four projects

OpenClaw: broad channels, native apps, one local gateway

OpenClaw is an open-source assistant that runs on your own computer and answers in the channels you already use: WhatsApp, Telegram, Slack, Discord, iMessage, Signal, Teams and more than 20 others. A single Gateway manages sessions, tools and channel connections, and the same setup can serve one person on a laptop or a team. Companion apps for macOS, iOS, Android, Windows and Linux add voice, Canvas, camera and screen actions. Models and agent harnesses (Claude, Codex, local models) are plugins you can swap, and its documentation states that by default it only checks for new versions daily, with anonymous statistics opt-in. It is MIT licensed and governed by an independent foundation.

Consider it if you want the widest channel and device coverage and are comfortable operating a Node.js service. Read the security guide first: tools run on the host unless you configure sandboxing, and inbound messages should be treated as untrusted.

nanobot: small, readable, Python

nanobot is an ultra-lightweight self-hosted agent framework written in Python, from the HKU Data Intelligence Lab. It runs in a browser WebUI, the terminal or chat apps (Telegram, Discord, Slack, WeChat, Feishu, email, Linear and more), with tools, long-term memory, MCP, scheduled automations and subagents around a deliberately small core. It works with OpenAI-compatible APIs and local models, with provider fallbacks, and exposes a Python SDK and an OpenAI-compatible API. MIT licensed, Python 3.11+.

Consider it if you want something you can read and adapt, or want to embed an assistant in your own Python project. It has fewer enterprise features than larger platforms, and the source install moves faster than the stable release.

CowAgent: the deepest coverage of Chinese platforms

CowAgent, formerly chatgpt-on-wechat, grew from a popular WeChat bot into a personal assistant and agent harness. One instance can serve a web console, WeChat, Feishu/Lark, DingTalk, WeCom, QQ, WeChat Official Accounts, Telegram, Slack and Discord at the same time, with text, images, files and voice on most channels. It has a Skill Hub, long-term memory and a personal knowledge base, agent teams, a web console for configuration and a one-line installer. MIT licensed.

Consider it if your users are on Chinese messaging platforms. The project states it is intended for technical research and learning, and that you are responsible for complying with local laws and each platform's rules, which matters for personal WeChat integrations.

Letta: memory first

Letta, formerly MemGPT, is about agents that keep memory and identity and improve over time by rewriting their own context. The active code now lives in the letta-code repository (the earlier V1 API server is retired): a terminal UI, desktop app, browser interface and messaging channels, with context tracked in git, skills, subagents, schedules and permission modes. Letta Cloud is the default backend, and some features such as remote computers and secrets require signing in. Apache-2.0 licensed.

Consider it if long-lived memory is the point. Because agents rewrite their own context, review what they have stored and what they are allowed to do from time to time.

Side by side

OpenClawnanobotCowAgentLetta
Language / runtimeNode.jsPython 3.11+PythonTypeScript (letta-code)
Best atBroad channels plus native device appsSmall, hackable coreChinese platformsPersistent, self-improving memory
InterfacesChat apps, Control UI, native appsWebUI, TUI, chat appsWeb console, many chat appsCLI, desktop, browser, chat apps
Hosted optionNone (self-hosted only)None listedOptional online trialLetta Cloud (default backend)
LicenceMITMITMITApache-2.0

Desktop-first alternatives

If you would rather open an application than message a background service, two other tools in the directory take a different approach. AionUi is a desktop "cowork" app that runs a built-in agent and also detects and manages CLI agents such as Claude Code and Codex in one interface. Cherry Studio is a desktop client for many model providers and 300+ preconfigured assistants. Neither is a background service you message from your phone, but both are easier to start with.

Safety checklist for any assistant that can act

  • Start with the narrowest workspace. Point the assistant at one folder, not your home directory.
  • Use approvals or sandboxes for shell commands and file writes until you trust the setup.
  • Treat inbound messages as untrusted. Anyone who can message the assistant can try to steer it; restrict who may talk to it and require pairing where offered.
  • Keep keys out of shared files and prefer scoped, revocable API keys.
  • Do not expose the control interface to the internet without authentication and TLS.
  • Review memory and skills periodically, especially for assistants that write their own.

Browse more projects in the Workflow Automation and Multi-Agent categories, and see our guide to choosing an MCP server before connecting tools to any of these assistants.