Changelog & monthly updates

Flo here — last month was about control: intelligence levels, on-demand usage, social scheduling. This month is about reach. Until now, Notis lived in the cloud — its own computer, its own skills. Today it breaks out.

Two features, one idea: meet you where the work actually happens. The first brings Notis onto your own machine — your shell, your local tools, the desktop apps that never had an API. The second goes the other way: it puts Notis inside the coding agents you already use — Codex, Claude Code, Cursor — so the skills and integrations you set up once follow you everywhere.

Let's get into it.

Your Computer

The Cloud Computer we shipped in v3 gave Notis a machine of its own. But some things only live on your machine — a script in a folder, a CLI you've configured, an app with no API, a tool running on localhost. So we gave Notis a way in.

With the desktop app, Notis can now act on your computer, three ways:

• Local Shell — run shell commands on your machine: open files, run scripts, drive the CLIs you already have set up.

• Local MCP — connect Notis to MCP servers running locally, either as a local command or on localhost. Your private, local tools become Notis tools.

• Desktop Use — hand Notis your mouse and keyboard to drive desktop apps that have no API. If you can click it, Notis can.

This is opt-in and it stays under your control. Notis always prefers the safe path first — a connected integration, a Composio tool, or the Cloud Computer — and only reaches for your machine when that's what the job needs. You set the approval level, from ask me every time all the way to run everything, and you can dial it back whenever.

Learn more about Your Computer.

Notis for Coding Agents

If you live in Codex, Claude Code, or Cursor, you know the tax: every new agent means re-teaching it your skills and re-connecting every integration from scratch. We think you should set that up once.

Connect Notis once, and your coding agents work through your Notis account:

• Sync your skills — author a skill a single time and keep it available in Codex, Claude Code, Cursor, and Notis. Update it in one place and the change follows everywhere.

• Share your integrations — Gmail, Slack, GitHub, Notion, Linear, Calendar: connect them once through Notis instead of wiring each agent up separately. Your agents get the access; you skip the setup.

• Search your memory — your agents can call search_memories straight from the CLI, so Codex or Claude Code can pull from everything Notis remembers about you and your work — no copy-pasting context, no re-explaining yourself.

It all runs through the Notis CLI: install the desktop app, authenticate, and your built-in skills sync straight into your local agents.

Learn more about Notis for coding agents.

Now running on GPT-5.6

Two features above — one upgrade underneath. In June we gave you the intelligence dial: Low, Medium, High, and Auto. This month we upgraded what sits behind it. Notis now runs on OpenAI's brand-new GPT-5.6:

• Low runs on GPT-5.6 Luna — quick and light for everyday work.

• Medium and High run on GPT-5.6 Sol, OpenAI's newest frontier model — dialing up the reasoning as the problem gets harder.

Nothing changes in how you use it: the dial is the same, and Auto still picks the right level for you. The difference is that every task now runs on a sharper, more capable model underneath.

🐞 Other improvements

• Usage, per thread: the Manager now shows how much each thread has used — so you can see exactly where your plan is going, right next to the work.

• Responsive follow-ups: send Notis another message while it's mid-task and it now picks it up right away, instead of making you wait for the first one to finish.

• Smoother trials: during a free trial, your plan details and the activate button now show up exactly where you'd expect them.

• Memories uploads, fixed: dropping files onto your Memories page works reliably again.

• A steadier month: plenty of behind-the-scenes fixes across voice, billing, and automations.

🗞️ In Other News

• We shipped a brand-new Integrations directory this month — browse everything you can plug into Notis and see how to automate more of your work.

• Two things that resonated this month: on LinkedIn, "they have 2,000 paying teams and raised $80M; we have 1,000 and raised $0 — does efficiency beat funding?", and on Instagram, a reel on ADHD, dopamine, and outsourcing the tasks you can't force yourself to do.

• I post daily across X, LinkedIn, and Instagram — come say hi.

📰 AI News

Here's what you might have missed in AI since June:

🔥 OpenAI ships GPT-5.6 in three tiers — and it trained itself — OpenAI released GPT-5.6 as Sol, Terra, and Luna ($5 / $2.50 / $1 per million input tokens), with Sol hitting 91.9% on Terminal-Bench 2.1 and 90.4% on BrowseComp, plus a million-token context window. It also folded Codex and ChatGPT Work into a single desktop "superapp." Most striking: OpenAI says Sol autonomously post-trained the smaller Luna model with "limited human hands on the wheel." (OpenAI)

🎙️ GPT-Live makes AI voice full-duplex — OpenAI launched GPT-Live and GPT-Live-1 mini, which listen and talk at the same time and quietly delegate hard reasoning to GPT-5.5 in the background. The new voice model scores 75.2% on BrowseComp versus just 0.7% for the old Advanced Voice Mode, and 84.2% on graduate-level GPQA. Live on iOS, Android, and web globally. (OpenAI)

🎭 Anthropic ships Sonnet 5 and brings Fable 5 back on a leash — Claude Sonnet 5 landed at the same $3/$15 pricing (promo $2/$10 through Aug 31) but benchmarks below Opus 4.8, signaling restraint over raw capability. Meanwhile Fable 5 — pulled June 12 over export controls — returned July 1 wrapped in a new safety classifier that blocks cyber-exploit jailbreaks with 99%+ effectiveness. (Anthropic)

⚖️ Apple drags OpenAI into federal court — Apple filed a trade-secrets suit in the Northern District of California, alleging OpenAI poached 400+ former Apple employees — including 24-year veteran Tang Tan, now OpenAI's chief hardware officer — and built its devices on stolen iPhone IP. Apple wants a jury trial and a forced redesign of OpenAI's hardware. (Bloomberg)

🧠 Meta can now read your mind — no implant required — Meta's Brain2Qwerty v2 decodes typed sentences straight from brain activity using an MEG scanner instead of surgical electrodes, hitting 61% word accuracy (78% for the best participant) — a leap from the ~8% earlier non-invasive methods managed. (Meta AI)

🛡️ The EU quietly greenlit "Chat Control 1.0" — The European Parliament let voluntary suspicionless scanning of private messages stand until 2028: the motion to reject fell short at 314 votes (361 needed). It lets platforms like Instagram, Discord, Snapchat, Gmail, and iCloud scan the private messages of roughly half a billion people without warrant or suspicion. (Patrick Breyer)

🏛️ South Korea bets $880 billion on the AI stack — President Lee Jae-myung, alongside Samsung and SK Hynix, committed at least $880B to semiconductors, data centers, physical AI, and robotics, framing it as national "survival." SK Hynix had already crossed a $1 trillion valuation in May. (BBC)

🦾 Europe's biggest robotics round yet — Germany's NEURA Robotics raised up to $1.4B in the largest Series C ever for full-stack robotics, at a ~$7B valuation, backed by Nvidia, Amazon, Qualcomm, and Bosch. With a $1B+ order book for its 4NE-1 humanoid, NEURA is aiming to ship millions of robots by 2030. (NEURA Robotics)

That's it for July. v3 gave Notis a computer and skills of its own — this month it steps onto yours, and into every agent you already run.

See you next month.

Flo

P.S.: Record a video testimonial (in any language) and get a free month of Notis on us.