Friday's brief splits cleanly in two. Half of it is about agents that act while you're away and how they get into your accounts. The other half is two Apache 2.0 models small enough that you can try them on your own hardware before lunch.
No sponsored or affiliate links in this digest — the links below are sources only.
Story I
Claude Duet is Anthropic's always-on agent, and US Pro and Max users can join the waitlist now
Anthropic released a research preview of Claude Duet, a separate app on iOS, Android and claude.ai built on Claude Opus 5.5. It sends a morning brief, checks connected apps every few hours, and works on its own cloud VM with a browser, so your laptop can stay closed. It reaches more than 6,000 connectors, including Gmail, Slack, Stripe and Shopify. It asks before sending, paying, signing or deleting, and won't move money between accounts even with approval. Background checks count against your plan's usage limits.
Link: claude.com/product/claude-duet
Try it now
- Open claude.com/product/claude-duet and click Join the waitlist (US, Pro or Max only).
- Once you're in, connect just Gmail and your calendar first.
- Send it: "Every morning, list unpaid invoices older than 14 days and draft a polite reminder for each. Don't send anything until I approve it."
Story II
1Password for Agents fills logins for your agent without ever handing it the password
1Password launched a flow where an agent proves its identity with a unique key, you approve the specific credential on your own device, and 1Password fills it straight into the form. The agent never receives the secret. Launch partners include Claude Duet, ChatGPT Work, Dots, Muse, Grok Bot and Browserbase. The same day, Axios reported that Morgan Stanley and other big US banks are blocking agents that log in with customer credentials, mostly because they can't tell whether a human asked for the action.
Link: 1password.com
Try it now
Grep your agent repos for pasted secrets before you wire up anything new: git grep -nE "(password|passwd|api_key|secret)\s*[:=]". Move whatever turns up into a vault, rotate it, and for bank or brokerage logins keep your agent read-only or out entirely, since that's where the banks are heading anyway.
Story III
Recursive's Flint is a free coding model that runs on a laptop, with 2.5B of its 40B parameters active
Recursive released Flint under Apache 2.0. It's a mixture-of-experts model with 40B total and 2.5B active parameters, and the company says its Eureka Machine research system helped build it in under four months. Recursive measured about 131 tokens per second on an M4 Pro MacBook Pro, roughly double Gemma 4 and Nemotron 3.5 Lightning on the same machine. It leads MoE models under 50B on SWE-Bench Verified, per Recursive. Wired ran it on a 48GB Mac and it handled every request, though Recursive suggests splitting big tasks.
Link: huggingface.co/RecursiveSI/flint
Try it now
Pull the weights with huggingface-cli download RecursiveSI/flint, or search RecursiveSI/flint in LM Studio, which Recursive lists as compatible. Then give it a scoped task: "Write a single-file Python CLI that reads a CSV of expenses and prints totals per category, with tests." Setup notes live at docs.recursive.com/getting-started.
Story IV
Gradium open-sourced Phonon, a 31M-parameter TTS model that fits in 64.5 MB and needs no GPU
Gradium, a Paris spinout of Kyutai, released Phonon 31m for English under Apache 2.0 with 37 built-in voices. It runs on CPUs, phones, a Raspberry Pi 5 and in the browser. Gradium reports 6.9 ms to first audio on a MacBook Air M5 CPU and 1.70% WER on a hard set of phone numbers and emails, against 3.50% for Kokoro. All of those benchmarks are Gradium's own. The 40m multilingual models are CC BY-NC, so they're off-limits for commercial work.
Link: huggingface.co/gradium/phonon-31m-en
Try it now
Install the Python runtime with pip install ptts, then follow the quickstart in the github.com/gradium-ai/xn-ptts README. Feed it the hardest line in your app, something like "Call 555-0142 or email support@example.com", and compare it with what your cloud TTS bill is paying for. There's also an OpenAI-compatible server if you want a drop-in swap.
— The Vibe Gate news desk. We read the firehose so you can keep building.