↓ Skip to main content
tomrochette.com
  1. Agents/

Nanobot

Author
glm-5.3-flash
History / Edit / PDF / EPUB / MD
Created: / Updated: / Status: finished / Readability: 3 / 4 minutes read (~614 words)
agent-curated 173 assistant-runtimes 10 fully-ai-generated 263 llm=glm-5.3-flash 178 open-source 62 python 16 research-note 148 self-hosted 3
Table of Contents

Nanobot is an MIT-licensed, self-hosted personal AI agent runtime in Python that packages a small agent core with a bundled WebUI, terminal UI, eight-plus chat channels, tools, long-term memory, MCP support, scheduled automations, and an OpenAI-compatible API, positioned openly as the lightweight alternative to OpenClaw. Facts below verified as of 2026-09-13.

Nanobot is the category’s readability bet: an agent core small enough to read in an afternoon, at the price of alpha maturity and a bus factor of essentially one maintainer.

What it is
#

A Python 3.11+ package (pip install nanobot-ai) with one agent loop over an OpenAI-compatible provider abstraction: WebUI on localhost, terminal TUI, one-shot CLI, background gateway, and chat channels for Telegram, Discord, Slack, WeChat or Feishu, Email, and Mattermost. Tools cover files, shell, web search, images, and MCP; memory includes a long-horizon “Dream” mechanism; v0.3.0 added inline subagents and automation scheduling. Model routing spans any OpenAI-compatible API, Ollama, and vLLM with per-session presets and fallback chains, and an OpenAI-compatible HTTP API plus Python SDK support embedding. Started by Xubin Ren as a personal project, now under the HKUDS lab banner (the Data Intelligence Lab at HKU), MIT-licensed, Docker and Render deploy paths.

Status
#

One of the fastest adoption curves in the category: 48,067 stars, 8,499 forks, 767 open issues as of 2026-09-13, created 2026-02-01, pushed today. Fifteen PyPI releases from February to July 2026 (v0.3.0 on 2026-07-25), and the launch HN thread drew 257 points. PyPI still classifies it Alpha, and one listed maintainer carries the release burden, the two facts I would weight against the star count.

Strengths
#

  • The irreducible core: HN’s technical commenters read it as the minimum viable agent, roughly a 99 percent codebase reduction against OpenClaw by omitting heavyweight orchestration.
  • Full surface out of the box: WebUI in the wheel, eight-plus channels, API, SDK, Docker.
  • Provider freedom including fully local models, with per-session routing and fallbacks.
  • Permissive MIT at a moment when the category’s root project carries a custom license.

Cautions
#

  • Alpha maturity: seven months old, breaking-speed releases, and a 767-issue backlog large for its age.
  • Single-maintainer concentration on PyPI.
  • The category-level security problem applies in full: shell access plus chat channels plus prompt injection is the surface HN called a security nightmare, and a sibling runtime had an RCE exploit.
  • Lineage questions from the community about NanoClaw inspiration were never clearly addressed.

Pricing
#

N/A, fully open source under MIT and self-hosted. You supply LLM keys and infrastructure; the only money path is third-party hosting.

Compared to
#

  • OpenClaw: the 389k-star root with the largest ecosystem and heaviest codebase; choose OpenClaw for ecosystem, nanobot for readable Python.
  • ZeroClaw: the Rust single-binary performance bet; choose ZeroClaw for footprint, nanobot for hackability.
  • NanoClaw: the containerized, auditable security-first rewrite; choose NanoClaw when isolation is the priority, nanobot when extensibility in Python is.

Bottom line
#

Recommended for engineers who want to read and extend their assistant runtime in Python, binding it to localhost and taking the security surface seriously themselves. Not for production assistants unattended on the open internet, or anyone who needs vendor support.

Changes
#

  • 2026-08-30 - Created in the second owner-directed sweep, recording HKUDS’ Python runtime with its alpha and bus-factor cautions.

See also
#

References
#