Nanobot

Most people who want a personal AI assistant face a choice between hosted products that own their data and heavy agent frameworks that demand real engineering effort to stand up. Nanobot, from HKUDS, tries to sit between those poles. It is a self-hostable personal AI agent runtime with a deliberately compact and readable core, shipped so you can run it as a single deployable unit rather than wiring together a stack of services. The pitch is ownership: you bring your own model keys, run the agent on your own machine or server, and keep memory, schedules and conversation history under your control. Rather than another orchestration platform, it presents itself as a lean runtime you can actually read, audit and extend without first learning a large and opinionated framework.

At its center is a simple agent loop: an incoming message triggers an LLM to decide whether to answer directly or call a tool, with memory and skills pulled in as context only when needed. This keeps the runtime light instead of layering heavy orchestration on every request. A provider abstraction lets it talk to OpenAI-compatible APIs, Anthropic, Gemini, DeepSeek and local models through Ollama or vLLM, with fallback models and prompt caching configurable. The distinctive part is reach across channels: Telegram, Discord, Slack, Teams, email and a built-in WebUI all feed one session context, and Model Context Protocol support means external tools plug in through a standard interface. Scheduling, reminders, web search and file reading come built in rather than as separate add-ons.

The clearest fit is an individual or small team that wants an always-available assistant they fully control, reachable from the chat apps they already use. Concrete scenarios the project highlights include scheduled monitoring tasks that run on a cron and report back, daily routine and reminder automation, document and knowledge management backed by persistent memory, and lightweight coding or research help. Because the codebase is small and the configuration is plain JSON, it suits developers who want to read the source, add a provider or a tool, and understand exactly what the agent is doing. Privacy-conscious users who refuse to route personal context through a hosted product, and tinkerers who enjoy self-hosting on a home server or VPS, are the natural audience here.

The honest caveats are real. Nanobot is pre-1.0, sitting around the early 0.x range, so the API and configuration format can change between releases and breakage on upgrade is plausible. Documentation reachability was inconsistent during our checks, with the project wiki returning errors, which raises the cost of getting started for anyone who is not comfortable reading code. The space is crowded with personal-agent and agent-OS projects, so its differentiation rests largely on the lightness and readability of the core rather than a unique capability, and that is something each team should evaluate against its own needs. The broad channel and provider matrix also means more surface area to keep working, and the project itself lists long-term memory and multi-step reasoning as areas still to improve.

Against commercial automation tools like Zapier AI or Make, Nanobot is a different proposition: those are polished, hosted and integration-rich but keep your workflows and data on their platform, while Nanobot trades that convenience for full ownership, a transparent codebase and no per-task pricing. Adopt it if you want a minimal, auditable agent you can host yourself and extend without committing to a heavy framework, and you are comfortable tracking a fast-moving pre-1.0 project. Hold off if you need guaranteed API stability, formal support and reliable documentation for a production enterprise deployment, or if you simply want point-and-click automation without touching configuration. For self-hosters and builders it is a credible, low-overhead starting point; for risk-averse production teams it is still early.