2026-07-17

Daily AI Digest: Claude signs in safely and local agents take on real work

This is the GolemWorkers daily AI digest, focused on practical agents that can complete useful work while people stay in control. Today, Claude gets safer access to logged-in websites, LM Studio brings private agent work to local models, and new releases extend agents into remote devices, shopping, security, and browser operations.

A woman approves a secure browser login with biometrics on her phone.

1Password lets Claude sign in without seeing your credentials

1Password for Claude lets Claude in Chrome complete browser tasks that require passwords and one-time codes, while the secrets stay out of the model and its memory. The Mac integration is available on business, family, and individual plans; each request shows which login is needed, requires the user's biometric approval, and grants access only for that task.

Source: 1Password
A designer reviews a presentation and spreadsheet produced from local documents.

LM Studio Bionic turns open models into a private work agent

LM Studio's new Bionic app can inspect and edit local code, organize files, research on the web, and create documents, decks, and spreadsheets inside a sandbox with previews and rollback checkpoints. People can run models locally for privacy, connect another machine through LM Link, or use larger open models in LM Studio's zero-data-retention cloud.

Source: LM Studio
A workshop owner speaks to a phone while a remote computer and camera node continue separate tasks.

OpenClaw adds mobile automations and more capable remote nodes

OpenClaw 2026.7.2 beta.2 adds mobile Automations, foreground Voice Wake on Android, and camera, location, and notification capabilities for headless Linux nodes. It also improves remote coding sessions across cloud workers and paired hosts, so owners can start, resume, and supervise agent work beyond their main computer.

Source: OpenClaw release notes
A customer checks a household delivery while confirming payment on her phone.

HSBC and Visa run a live agent-assisted card purchase

HSBC UK says one of its cards initiated an end-to-end agentic transaction on a live merchant website with Visa. The companies are working toward AI services that can find and buy travel or household goods, with biometric authentication, explicit permissions, and payment safeguards keeping the customer in control.

Source: HSBC UK
Security engineers trace a branching attack path to one verified code flaw.

Capital One releases an agent that hunts and fixes exploitable code flaws

Capital One's open-source VulnHunter uses three Claude Code agents to trace attacker-accessible paths, challenge its own findings, write a security test and focused fix, and independently verify the repair. Authorized security teams can run the Apache-licensed toolkit on individual repositories or unattended batches, with confirmed findings turned into reviewable issues and pull requests.

Source: Capital One
An engineer reviews a failed browser workflow beside a proposed code repair.

Libretto opens a repair pull request when browser automation breaks

Libretto's free Playwright PR Agent starts only after an existing browser script fails, inspects the live page, and opens a GitHub pull request with a proposed code change. Teams can add it to current Playwright jobs without replacing their runtime or browser provider, then keep their own retry and fallback rules for the failed run while a human reviews the repair.

Source: Libretto

Agent idea of the day

Catch subscription price hikes before they renew

A couple reviews subscription renewal dates, bills, and an approval summary at a kitchen table.

What this agent does

Monitor recurring services for renewal notices and price changes, then prepare one approval brief before money moves.

Best for: Households and small teams with several software, media, insurance, or membership renewals to track.

Give it

  • Read-only access to a dedicated folder of receipts and renewal emails
  • A list of approved vendors, current prices, and any must-keep services
  • Calendar access and, optionally, human-approved 1Password access for account-page checks

Tell it to

  1. Search the approved email folder for renewal, trial-ending, and price-change notices.
  2. Match every notice to the vendor list and calculate the new annual cost and the decision deadline.
  3. Open an account page only when evidence is missing, and wait for my biometric approval before using a login.
  4. Flag duplicates, unused services, material price increases, and cancellation deadlines with a link to the evidence.
  5. Draft a weekly approval brief that recommends keep, negotiate, downgrade, or cancel, but take no action until I decide.

Run it: Run every Monday morning and immediately when a renewal, trial-ending, or price-change email arrives.

You get

A concise renewal brief with vendor, old and new price, annual impact, deadline, evidence link, recommendation, and a clear approval request.

Keep a human in control

  • Never purchase, renew, downgrade, or cancel a service without explicit human approval.
  • Never expose passwords, one-time codes, full payment-card details, or unrelated email content to the model.
  • Check only approved vendors and use read-only access wherever possible.
  • Show the original notice or account evidence for every price and deadline.

Feasibility: 1Password documents that Claude can complete logged-in browser tasks while passwords and one-time codes stay outside the model, with a user-consented biometric approval for each credential request and task-scoped access. Source: 1Password for Claude →

GolemWorkers — OpenClaw ready to work: launch in seconds, no setup, online 24/7.