Atomic Agent v0.5.4 is out! ✦ Hugging Face and Ollama models ✦ Coding mode: plan, auto, or ask-first ✦ Context window: set the size of memory ✦ Fully redesigned TUI: scroll, click, select, copy ✦ Import data from Hermes, OpenClaw or ClaudeCode Built for local models:

How long I have been waiting for THIS 🦋👍👍

Same, honestly. Grab it and tell us what feels off, we ship fast:

I'm excited about this release! Looks quite interesting. Great work!

Thank you. If you try it, the thing to poke at is the context control, you decide how much history goes into each request. Discord is where we collect the feedback:

Best TUI on the net 🪽

ngl we spent way too long on that

a clickable TUI. finally someone remembered that terminal people own a mouse too. thank you friends!

huge release!

Appreciate it. Go break it and tell us where it hurts:

WOW incredible TUI! Great job!

Thanks Ivan, that means a lot. The tricky bit was that mouse reporting and native drag select are mutually exclusive in a terminal, so a drag now pauses mouse mode for 10 seconds instead of asking you to hold a modifier:

redesigned TUI looks pretty clean

Thanks. The look is half of it, the other half is that it answers the mouse now, scroll, click, drag to select, copy. Full notes:

Mouse support is a nice touch.

Impressive as always hope that atomic build us the first advanced local live chat that utilize local llms it will be futuristic

Agreed on the default, asking first should never be a setting you have to find. Ours is four modes on one key, from read only planning to full auto, and the model stays yours either way:

The four modes sound great, but I was specifically hyping that voice + ai combo 😊 By live chat I meant voice conversation, like an actual back-and-forth talk. But it doesn't have to be inside the agent itself, might be a tool of its own. Keep up the great work! 🚀

the ClaudeCode import is the interesting bit here — does it preserve tool history and project context, or just the messages?

More than messages. Sessions come over as full turns from the project jsonl, plus your skills, your CLAUDE.md and auto memory notes, and MCP servers read from claude.json. Onboarding takes the 100 newest transcripts, the CLI takes everything:

This is a solid upgrade for local AI workflows. The redesigned TUI and flexible coding modes stand out.

Appreciate the read. Modes plus the context control are meant to work together, keep it in plan while it explores, then let it write. What is your local setup?

Thanks ⚘️

This is have web version? Like opencode web

No web version right now, it is terminal plus CLI, and you can drive it from Telegram when you are away from the machine. Worth an issue if you want it tracked:

this looks intriguing, congrats on launch!

Thanks. It is open source, so poke around the code before you even install:

Nice, Congrats 👏🏻

Thanks. Give it a spin and tell us what breaks:

huge congrats team

Appreciate it. Come say hi in Discord, half the roadmap comes from there:

Good work

Thanks. It is all open, have a look under the hood:

Coding mode plus a hard context budget is the right pair. The silent failure mode is still fat tool descriptions eating the window before the repo map does. Audit tools the same way you audit memory size.

Fair hit. The scaffold line in our context panel is where that lives and it is 5.9k before you type anything, so it is at least visible rather than silent. No per tool budget yet though, only the skill catalog has one. Worth an issue:

local-first + ask-first coding mode is the right default. heliox is MIT on-device desktop agent: 5 permission tiers, 157 actions, files stay on the box.

Best TUI fr

Glad it lands. Collapsible sidebar and clickable tool cards were the two we fought hardest for, more coming: