Breezell

Light as a breeze. Made for developers on the long coding journey. Building Breezell, an AI code editor.

[Breezell @BreezellCode]

@mayuri_3015 Dioxus

[Breezell @BreezellCode]

@pcshipp All in! All in!

[Breezell @BreezellCode]

I feel like AI has actually made things worse. Whether it's open-source or private projects, people seem to constantly look for the easiest way to bypass testing. Or they'll let AI generate a few lines of code and then spam authors with PRs without even verifying the architectural integrity or checking if it fits the existing structure.

[Breezell @BreezellCode]

@sahildarz I totally get how you feel. We only have two people on the team, and after releasing version 1.3.4 yesterday, opening a new project resulted in a black screen. We had to pull an emergency overtime shift for two hours just to rebuild the package—it was an absolute nightmare!💤

[Breezell @BreezellCode]

@azureraves @thsottiaux @aaryankushwah Thanks, but we currently have no plans to open source it.

[Breezell @BreezellCode]

@thsottiaux @aaryankushwah @thsottiaux I can’t DM you from this account. I messaged you about this a few days ago. What’s the right way for a small team to request Sign in with ChatGPT approval?

[Breezell @BreezellCode]

@LazyIDE Fully managed

[Breezell @BreezellCode]

We've been getting a lot of questions lately about multi-model, multi-agent orchestration.
Breezell already ships Teams and subagents, and both let you pin the model they run on. By default, whether a subagent gets spun up is decided by task complexity; it's only force-invoked when you explicitly ask for it.
We deliberately did not build a system where multiple models and multiple agents automatically pick each other and orchestrate each other. It's not that we couldn't. Every extra handoff in an orchestration burns another round of tokens, and errors propagate downstream. Most of the time it's neither faster nor better. Mixing models can also introduce unexpected failure modes: divergent styles, context understanding, and tool-use conventions make handoffs more likely to fall out of sync. Light usage is fine — Claude for architecture, GPT for implementation can occasionally produce positive results — but it doesn't belong as the default architecture. Subagents are a good fit for work that can run independently in parallel and would otherwise pollute the main thread; not every scenario deserves one. At best, this is a token-saving strategy under hybrid routing. So our default is routing, not orchestration: first decide whether decomposition is worth it, and if not, just get the job done in a single thread. The goal is less pointless analysis and fewer wasted tokens.
Team is carved out as a separate, explicit option so that the developer — not Breezell — gets to judge whether a scenario actually warrants a Team. We won't auto-spin a team up for you.
Don't sacrifice basic convenience for the sake of elegance.
In practice, orchestration is just one AI writing prompts for another AI. It will make mistakes. Even with human review, you save no time — you've merely shifted the cost from writing prompts to reviewing them. You might as well just write the prompts yourself.
This isn't a rejection of multi-agent collaboration; it's just that for coding specifically, it may not be the right fit yet.
We can't even reliably judge whether AI-written code is correct — so how can we be sure an AI can get acceptance right?
We can't even verify whether the orchestrator understood the requirements correctly. The orchestrator has to produce the "role cards" — in one shot, or incrementally over multiple passes — and may even keep expanding them, propagating its errors to every other role and causing needless contamination. Where this goes next still needs to be validated.
#AICoding #VibeCoding #AIAgents #Codex #ClaudeCode

[Breezell @BreezellCode]

@anrayama Exactly. This is critical since most major vendors don't surface cache metrics natively. You basically have to sit a proxy in front of them just to see if caching is actually working.

[Breezell @BreezellCode]

We’re building Box Mode in Breezell — one console box for every model you call.
Cache hit rate, LLM volume, token burn, and estimated savings sit in one row. Below: a yearly activity heatmap and a 30-day hit-rate curve. On the side: providers stacked — OpenAI-compatible, Codex, OpenRouter, Poolside, xAI Grok — with rounds, cache vs input, and hit %.
Filter conversations and providers. Copy a summary. Stay in flow.
This will be the best visualization mode for VS Code secondary-development programs.
#Breezell #VSCode #AICoding #DevTools #LLMOps #BuildInPublic #OpenRouter #Grok #Codex

[Breezell @BreezellCode]

@sahildarz Context handling is crucial in agent development because it can lead to AI misjudging the objective. Most agent tools have committed this error.

[Breezell @BreezellCode]

@sahildarz Thank you for your suggestion.

[Breezell @BreezellCode]

@jilyannori Thanks!

[Breezell @BreezellCode]

1.3.3 · 2026-09-07
New Features
Box workbench
The title bar now switches between Box and Agent in one click. Box is a standalone management surface that brings together models and providers, account and usage, Super Memory, Code Review, the MCP marketplace, the database browser, data cleanup, and Settings.
Models: Provider config sits next to the model list. Add or remove OpenAI-compatible endpoints. Choose a network path per provider — including official Breezell models — Direct, System proxy, Global proxy, or a custom proxy. Probe and test connectivity in one click.
Vault: Encrypts relay endpoints and account credentials, then fills them in where you need them. No more copying keys around.
Code completion (FIM): Completions can use their own connection and each vendor’s native FIM protocol. Major vendors ship as presets.
Data cleanup: Inspect local data stats, compact the database, and run integrity checks. Every maintenance pass is logged.
Database browser: Rebuilt, with multiple connections and clearer grouped navigation.
MCP marketplace: Organized by category, with a redesigned page.
Privacy notes now use a card layout and cover more of the details.
Breezell Mind
Conversations keep a running working notebook. When a long thread hits the budget line, older history is folded down to the notes and the essentials.
Quote this conversation is lighter: quoting another thread only carries the latest turn plus the files involved. Full history is read on demand.
Context compression now prepares and refreshes a background handoff document. The raw record stays on disk. You get a warning if compression has run too many times.
Conversation
After a turn finishes, you get a few “what you can do next” suggestions — click to continue.
A conversation can open as an editor tab, side by side with the code.
@ mentions for files and symbols are as fast as Quick Open. / opens the Skills menu.
Model menu: Breezell Max Plus toggle.
Agent system prompt rewritten: clearer about what each turn is for. After real work lands, you get a short wrap-up. Comment language follows yours.
File paths in messages render as clickable file chips and stay consistent across renders.
Models
Added GPT-6 Astra, Claude Fable 5.1, Gemini 3.8 Flash, Muse Spark 1.3, and OpenCode Go Omen Alpha.
Models without built-in web search now fall back to the provider’s native search.
Claude Sonnet 5 pricing updated to the latest official rates.
Editor
Edit Markdown in place and toggle preview without leaving the file.
Diff view adds a Diff View submenu for comparison modes. Custom editors can show diffs.
When opening a file, you can choose among editor types.
Feature options in Settings redesigned. Git commit-message generation has its own model card.
Account panel restyled as a membership card: plan badge, balance, expiry, and Super Memory quota at a glance. The account button shows your avatar.
Fixes
An unclosed < in a stream no longer freezes the rest of the text. Reasoning-pane code renders as plain text; code fences without a language still highlight.
Empty model replies trigger an automatic retry instead of silently ending the turn. Streams the upstream later marks empty are kept.
Token counts after context compression now match what is actually sent.
Changing the conversation model no longer overwrites the models chosen for Apply and Git commit messages.
Reasoning auto-scroll no longer detaches on its own; streaming code blocks no longer flash white.
Analyze sub-agents no longer spuriously ask you to switch to Agent mode.
The built-in browser no longer types keystrokes into the AI input when incognito is turned off.
Poolside Laguna requests that include a screenshot no longer fail the whole call; they now go through the vision proxy.
Attachment token estimates before send are accurate. Emergency compression can fold the current turn.
Error cards name the provider that actually failed.
Terminal extension commands no longer wait out the full shell-integration timeout before they run.
Plan-execution bubbles keep their corner radius. User-message shadows no longer cover the text below.
Model hover settings align to the row they belong to.
The code-index page in Settings no longer rescans the whole workspace on every open. Build artifacts are excluded from indexing.
Linux: Restored the 8 KiB Node buffer pool to avoid regressions in some workloads.
Improvements
repo_map no longer freezes the UI for tens of seconds. Line-level diffs for streaming edits and conversation persistence both leave the UI thread.
Explore tools have timeouts — a dead network will not pin the Agent for five minutes. Failed explores collapse to one quiet gray “tried” line.
Sidebar is lighter: MCP marketplace and Super Memory live in Box. History is paginated and grouped. Account badge placement tweaked.
Runtime upgraded to Electron 43.6 (Chromium 150 / Node 24).
Installers ship only the binaries for the current platform, so packages are smaller.

#Breezell #AICoding #VSCode #MCP #BuildInPublic

[Breezell @BreezellCode]

@kimmonismus Welp, guess I have to get out of bed and start working now. Damn it.

[Breezell @BreezellCode]

@claudeai My quota is about to go up in flames like a rocket launch. 🚀

[Breezell @BreezellCode]

@claudeai The GOAT of coding! Claude always feels so genuine.

[Breezell @BreezellCode]

We built the pricing page before we built the billing system. 😂

Very professional of us.

Custom API.
Trial.
Pro.
Pro+.
Ultra.
Teams.
Enterprise.
Everything looks ready.
Except for one tiny detail:
We currently can’t actually charge you. 💀
To be clear: we’re not giving away free API usage.
The good news: bringing your own API key to Breezell is ridiculously simple.
Add your API key, connect it, and you’re good to go.

I genuinely think this might be the simplest BYOK setup in the world.

No complicated setup. No unnecessary steps. Just bring your key and start coding.

The paid plans?

They’re patiently waiting for us to finish the part where money changes hands.

🌐 https://t.co/MxoBXlCKKk
📚 https://t.co/UeyJezraRt

#BuildInPublic #AIcoding

[Breezell @BreezellCode]

We’ve been quietly building Breezell.

If you haven’t heard of it yet: Breezell is an AI-powered code editor built on Code – OSS, designed to bring Agents, multiple AI models, deep codebase understanding, memory, skills, browser tools, and parallel AI teammates into one workspace.

And today, Breezell 1.3.2 is here. 🚀

This release is heavily focused on speed, stability, and long-running AI workflows:

• Long chats stay responsive while generating
• Smarter context overflow handling + local history recovery
• Faster Code Index that respects .gitignore
• Much smoother large code blocks and Diffs
• Fixes for WSL / Remote, Agents, browser, streaming, memory, and workspace state
• Separate model selection for Git commit messages
• GPT-5.6 reasoning levels, including Pro
• Hunyuan HY4, Kimi K3, Qwen3.8 Flash, GLM-5.3 Flash & more
• Grok Imagine 2.0 image generation + 1.5 video generation
And a ridiculous number of smaller bugs are gone.
Breezell is still small, but we’re shipping fast.
If you build software with AI, I’d love for you to try it and tell us what you think. 👀
I think you’ll like it. ❤️
🌐 https://t.co/MxoBXlCKKk
📚 https://t.co/UeyJezraRt
#AIcoding #BuildInPublic

[Breezell @BreezellCode]

@thsottiaux Happy Thursday!