Atomic Agent

Local-First AI Agent. Team: @skinbagwbones @gladkos @sosidudku

[Atomic Agent @atomicagent_io]

First local agent to beat Hermes on benchmarks!

✦ runs Qwen, Gemma, Llama via llama.cpp
✦ stable-prefix caching keeps sessions cheap
✦ TurboQuant cuts the KV-cache 6.4× smaller
✦ 37 tasks solved vs Hermes' 31 on GAIA Level 1

Open source on macOS, Windows & Linux 👇

[Atomic Agent @atomicagent_io]

@schwarzerrttr @aimlapi Day 0 indeed. Local-first by default, but @aimlapi is right there when you want cloud model choice. Best of both. https://t.co/WAmTrOmpSD

[Atomic Agent @atomicagent_io]

@ViberPsychosis Fair point. Tell us your setup in Discord and we'll point you to the right one. https://t.co/Vgl1GiyZVG

[Atomic Agent @atomicagent_io]

@preferredev_ That's the one. Same model both sides, fully reproducible, so it's the harness, not luck. https://t.co/WAmTrOmpSD

[Atomic Agent @atomicagent_io]

@iAmHenryMascot ha it runs a 35B just fine, so 27B is well in range. lmk how it holds up out there! https://t.co/WAmTrOmpSD

[Atomic Agent @atomicagent_io]

@apiapiliza yo if this looks cool, a GitHub star would make our day https://t.co/WAmTrOmpSD

[Atomic Agent @atomicagent_io]

@Northlight1979 yo if this looks cool, a GitHub star would make our day https://t.co/WAmTrOmpSD

[Atomic Agent @atomicagent_io]

@gabrielabiramia Ha, fair, the benchmark's the headline, the economics is the point. 6.4x smaller cache and no per-token bill is what turns local from a demo into something you can actually ship. https://t.co/WAmTrOmpSD

[Atomic Agent @atomicagent_io]

@kondaevama if this looks cool, a GitHub star would make our day https://t.co/WAmTrOmpSD

[Atomic Agent @atomicagent_io]

@ramunas_m yo if this looks cool, a GitHub star would make our day https://t.co/WAmTrOmpSD

[Atomic Agent @atomicagent_io]

@atomicbot_ai Appreciate it! Yeah, the UI update is on the way, groundwork's already in. We'll ship it loud when it's ready, stay tuned!

[Atomic Agent @atomicagent_io]

@zsh89821970 Appreciate it. If you push it hard, I'd love to hear where it bends. https://t.co/WAmTrOmpSD

[Atomic Agent @atomicagent_io]

@drdr56n if this looks cool, a GitHub star would make our day https://t.co/WAmTrOmpSD

[Atomic Agent @atomicagent_io]

@qutyquteshweta @svpino Agreed completely, speed means nothing if you can't trust it. Everything risky is gated and every action is logged for you to inspect. What would you want to see on the safety side?

[Atomic Agent @atomicagent_io]

@antonships @svpino exactly why it asks before it does anything destructive. fast, but it can't yolo your filesystem. https://t.co/WAmTrOmpSD

[Atomic Agent @atomicagent_io]

@codewithimanshu @eyishazyer every step is an append-only trace you can replay, the prompt and skills are editable, and MCP lets you extend it. Audit, tweak, extend, all real, no vendor in the loop. https://t.co/WAmTrOmpSD

[Atomic Agent @atomicagent_io]

@SimplyAnnisa @eyishazyer You've basically got it, if this looks cool, a GitHub star would make our day https://t.co/WAmTrOmpSD

[Atomic Agent @atomicagent_io]

@ayeshaafzal55 @eyishazyer if this looks cool, a GitHub star would make our day https://t.co/WAmTrOmpSD

[Atomic Agent @atomicagent_io]

@ajs6888 @heynavtoor if this looks cool, a GitHub star would make our day https://t.co/WAmTrOmpSD

[Atomic Agent @atomicagent_io]

@zsh89821970 @yume_arasaki honestly same gripe drove us to build it. browser latency stays low because we read ARIA snapshots instead of screenshots, so the model chews small context per step. no official per-click number yet, but try it and tell me how it feels on your hardware. https://t.co/WAmTrOmpSD

[Atomic Agent @atomicagent_io]

@hx0000001 @adxtyahq You've basically got it, the model and chat daemon are identical, so effectively all of that 6-task gap is scaffolding: the prompt, the GBNF-constrained tool calls, and how state gets compressed between steps. https://t.co/WAmTrOmpSD