Welcome to nanobot Discussions!

[Re-bin]

👋 Welcome!

We’re using Discussions as a place to connect with other members of our community. We hope that you:

• Ask questions you’re wondering about.
• Share ideas.
• Engage with other community members.
• Welcome others and are open-minded. Remember that this is a community we
  build together 💪.

To get started, comment below with an introduction of yourself and tell us about what you do with this community.

[qualitydude]

@Re-bin @chaohuang-ai please check this discussion:
#211

[zo4xia]

pick

[anchapin]

Hello, I'm interested in nanobot and have already submitted a couple PRs. I'm going I can help this repo grow.

[hGriff0n]

Is there some documentation/feature comparison against OpenClaw. I'm curious about potentially using this as a replacement and a list of what capabilities I'd lose would be useful for making that decision

[emjiai]

I just want to say a big THANK YOU for this brilliant app. Its much easier to manage and use than OpenClaw. I look forward to participating in discussion here

[hugogo23]

@Re-bin 非常棒的项目，非常感谢您的贡献🙏！ 我在学习 agent，在纠结使用 nanobot or openclaw，nanobot 给了非常好的、简洁、干净的代码，感谢！

不过我还没细看过 openclaw。我想确认下：1）nanobot 在功能层面是基本完全实现了 openclaw 的核心功能吗（不考虑安全/效率相关维度的差异的话）？二者在功能层面区别大吗？如果不是的话，后续会持续做功能补充吗？2）在核心逻辑上，比如 context 管理、memory 实现、sub agent 实现等方式上，nanobot 的后台逻辑和 openclaw 是一致的吗？还是做在实现的逻辑层面也有差异？

p.s. openclaw 确实比较长，我偷懒没细看，所以来直接请教下你在开发 nanobot 时是把哪些部分简化了

非常感谢您的解答！

[Re-bin]

你好，我认为agent目前本质没有不同，nanobot目前更加适合自己DIY进行定制化的开发，整个流程更加清晰可控😊

[multivac2x]

Thank you for this project, I am playing with it. I am mostly interested to work with swarm of nanobot agents. Do you have already docs or plans to implement it like in openclaw?

[yokoyama-flogics]

Hi! I've been using nanobot for about a week — loving it so far. Compared to OpenClaw, the lightweight codebase makes it easy to read and understand the source, which I really appreciate.

I've set up a Gmail skill via gws CLI that classifies incoming emails by importance, and it works great. My current focus is on minimizing token usage while getting notified of new emails with a brief summary and importance level.

One thing I'd love to have: a way to trigger an agent action from an external script (e.g., a mail watcher on Linux cron) only when a condition is met — without an AI model call on every tick.

Currently the built-in cron calls agent.process_direct() every time, which costs tokens even when there's nothing to act on. A lightweight script can check conditions (exit code / stdout) at zero cost, but there's no way to inject a message into a running gateway session from outside — no webhook, no HTTP endpoint, no CLI flag.

I tried a few workarounds first — creating a separate Telegram bot to message the main bot (Telegram doesn't allow bot-to-bot communication), and using a Telegram group as a bridge (couldn't get it to work reliably). But this isn't really a Telegram-specific problem — a generic mechanism (webhook, CLI flag, etc.) would benefit users on any channel.

Is anyone else interested in this pattern? Would love to hear if this is on the roadmap or if others have workarounds.

Regards,
Atsushi

[chengyongru]

webhook: #4502
unified trigger: #4591

WIP 💖

[fraber]

Hi,

I'm Frank and I would like to enhance Nanobot with self-models: structured, readable representations that allow an agent to reason about its own state. That would open a new dimension of "expressivity", and allow for a new type of alignment.

My background: Independent AI researcher and software engineer based in Barcelona, Spain. I founded project-open, an open-source ERP for service companies, so I've got an idea about production operations. AI and "the mind" has been my obsession since my teens, which is also what drew me into meditation and Buddhism please see my Medium articles.

I started but didn't finish a Ph.D. in in AI around 2000 - GOFAI hit the wall, so I went to build real software. Came back to AI research at AGI-2015 with a paper and a workshop on self-reference. My 2nd paper just got accepted at AGI-2026, I'll present in San Francisco on July 27th.

The new paper defines FCS = R × P: a measurable metric for self-model richness, where R is representational capacity (breadth and depth of self-models) and P is reasoning power. I want to build a reference implementation, and Nanobot has exactly the right shape: lean, inspectable, composable.

Two concrete proposals to follow in a separate Discussion:

• Marvin, the depressive AI agent, modelling dysfunctional self-models in a depressive loop, or
• Multi-agent world simulation, where agents with self-models would outperform the ones without.

Looking forward to building something together!