Antonio Duarte

Building Maestrly · founder
AI agents that actually get the work done

[Antonio Duarte @antonioducs]

I'm Antonio Duarte, building Maestrly.

Local-first AI workspace: conversations, Maestro orchestration, Git worktrees, terminals, and project context — on your machine, with your providers.

I'll post real product updates here as I ship.
→ https://t.co/59aAj6qT8p

[Antonio Duarte @antonioducs]

@banhosdev foi em que tipo de tarefa? aqui a diferença apareceu mais em ui e feature nova, em bug chato ainda mando o codex dar uma segunda olhada

[Antonio Duarte @antonioducs]

@brunogsantos comecei no terminal tbm, o que me fez sair foi revisar diff, com 3 agentes rodando eu acabava aprovando sem ler direito
hoje cada um fica na sua worktree e eu reviso numa janela, mas entendo quem prefere cli pela velocidade

[Antonio Duarte @antonioducs]

@krishnanrohit makes sense, they spent the last decade reviewing code they didn't write, that's exactly the job now

[Antonio Duarte @antonioducs]

@keviduk the remove it before you send part is what makes this work for me, dictation straight into an agent always felt risky
do you keep the raw transcript around in case the rewrite drops something i said?

[Antonio Duarte @antonioducs]

Tell one model once. Every model remembers.

I told a Claude chat I always use pnpm, then opened a GPT chat on a different provider, and it already knew.

Maestrly's personal memory stays on your machine, and you can see, edit or delete every entry.

[Antonio Duarte @antonioducs]

@slipgatecentral 10 days non-stop with both is wild. how did you split them, one builds and the other reviews, or just whoever had quota left? baked lightmaps holding up on a phone is the part i didn't expect

[Antonio Duarte @antonioducs]

@dani_avila7 tool results is where the tokens always go for me. one big file read early and you pay for it on every turn after. would love a per-turn delta view to spot that moment

[Antonio Duarte @antonioducs]

@gabrielbuzziv os dois, mas o diff eu não leio sozinho: peço pra outro modelo revisar antes. ele aponta 2 ou 3 trechos e é ali que eu leio com calma. teste prova que funciona, não que o resto tá certo

[Antonio Duarte @antonioducs]

@brunofaggion concordo, o "code" no nome assusta mais que a tela preta. quem não é dev nem cogita abrir

[Antonio Duarte @antonioducs]

one task, three providers, one conversation

Maestro mode gives the orchestrator an agent pool, and each agent runs on its own model. here gpt builds the ui, claude sonnet writes the tests and gemini reviews

it delegates in parallel and waits for everyone before answering

[Antonio Duarte @antonioducs]

your bot runs on a server. your code lives on your laptop

now each computer can give a Maestrly bot access: pick the projects, the models and how far it can go. it starts the work here, in its own worktree, with its name on it

commands still wait for you

[Antonio Duarte @antonioducs]

"should work now" isn't a test

Maestrly agents can open what they just built in the built-in browser, click through it and read the console before telling you it's done

same browser you see in the conversation, so you can take over anytime

[Antonio Duarte @antonioducs]

heads up: artifacts now live only on your bot server. ones from earlier versions can be moved there or deleted from Settings → Artifacts

free, open source, macOS/Windows/Linux
https://t.co/XzeSpw1ZIy

[Antonio Duarte @antonioducs]

Maestrly v0.14 is out

- pop any chat into its own window and keep a few side by side, even on another monitor
- personal memory shared across your chats, even across providers. stays on your machine
- send a plan from a standalone chat to a project, on a new branch

[Antonio Duarte @antonioducs]

also new:

- Maestrly tools on by default, with a switch per group (terminal, browser, memory…)
- bots get one Linux desktop each, browser + terminal + files together (experimental)
- attach files to bot messages and download what they generate

[Antonio Duarte @antonioducs]

one message, three agents working in parallel.

ask Maestrly to open a conversation per task and it starts them all. each one gets its own git worktree and branch, so nobody steps on anybody's files.

you can even pick the model per task: Opus 5.5 here, GPT-6.1 there

[Antonio Duarte @antonioducs]

free, open source, macOS/Windows/Linux

https://t.co/XzeSpw1ZIy

[Antonio Duarte @antonioducs]

Maestrly v0.13 is out

new:
- use your Google AI Pro/Ultra plan (Gemini) through Google's official Antigravity server
- local dictation with whisper.cpp
- agents publish web pages as artifacts you can share and get comments on

[Antonio Duarte @antonioducs]

also in v0.12 + v0.13:

- agent activity folded into one line per answer (expanded view is one setting away)
- bots update Claude Code and Codex on their own
- artifacts can live on your bot server, so shared pages stay up while your laptop sleeps

[Antonio Duarte @antonioducs]

plan with one model, build with another. same chat, same context.

in Maestrly you can ask GPT-6.1 for a plan, switch to Claude Opus 5.5 mid-conversation, and it picks up with the whole history: edits the code, runs the tests.

open source, local-first, your own accounts