1Password Lets Claude Use Stored Credentials Without Seeing Them

A new "zero-exposure" browser integration lets Anthropic's Claude agent log into sites on your behalf while keeping passwords and MFA codes hidden from the AI.

1Password has launched a browser integration that lets Anthropic’s Claude chatbot use stored login credentials to complete multi-step tasks — without the AI model ever seeing the actual passwords or one-time codes [1].

For developers and businesses already using Claude for browser-based automation, the practical gap has been authentication: agents hit a login wall and either stall or require a human to step in. This integration is designed to close that gap while keeping credential data out of the model’s reach [1].

How the zero-exposure framework works

1Password built what it calls a “zero-exposure security framework” that injects credentials into a secure channel the Claude agent cannot read directly [1]. The AI is granted permission to use stored details for a specific task, but it cannot view passwords or multi-factor authentication (MFA) one-time codes [1].

Access is scoped tightly. According to 1Password, Claude is granted credentials on a per-task basis, and users approve or deny each request with a single biometric prompt [1]. After every autofill action, 1Password scans the page to confirm nothing in form submissions remains exposed before handing control back to Claude [1].

“The moment an AI agent takes control of the browser, 1Password locks down automatically, limiting access to only the credentials explicitly granted for the current task,” the company said in its press release. “Nothing else in the 1Password vault is reachable.” [1]

Availability and current limits

1Password for Claude is available now for Mac users across business, family, and individual plans [1]. Using it requires the 1Password desktop app and browser extensions, plus the Claude desktop app and browser extensions [1].

At launch, the integration appears limited to login-related credentials [1]. 1Password says support for payment cards and identity details — both of which can be stored in its vaults — will arrive sometime after the initial release [1].


Sources

  1. The Verge — Claude can now use your 1Password credentials for you

This article was drafted with AI from the cited sources and checked against them before publication. Spot an error? Let us know.

Anthropic's coding agent can now open, read, click, and type on external websites directly inside the app, with safety classifiers screening write actions.

Anthropic has added an integrated browser to Claude Code, letting the AI agent open, read, click, and type on live external websites without leaving the app [1].

For developers, the practical upside is direct access to documentation sites, issue trackers, and similar resources during a coding session — no more copy-pasting content between a separate browser and the agent [1].

How it works

The browser behaves like a standard tab-based browser and opens via a keyboard shortcut, according to The Decoder’s reporting [1]. Under the hood, Claude uses the same tooling it already relies on to preview local apps, extended with additional safety checks for external sites [1].

Classifiers screen any write actions — such as form submissions — before they execute [1]. Anthropic said Claude will not purchase anything, create accounts, or bypass CAPTCHAs (the challenge-response tests used to block automated access) without explicit user consent [1]. The browser also runs on a clean profile with no saved login credentials, limiting the blast radius if something goes wrong [1].

Controls for organizations

Anthropic announced on X that organizations can restrict the browser to an allowlist of approved external sites, or disable the browser tools entirely [1]. Developers who need Claude to act within their own authenticated sessions — where saved logins matter — are directed to use the Chrome extension instead [1].


Sources

  1. The Decoder — Claude Code now has a built-in browser that lets the AI read, click, and type on external websites

This article was drafted with AI from the cited sources and checked against them before publication. Spot an error? Let us know.

Agentic workflows can consume 700x more tokens per user query than a chatbot, erasing the savings from falling model prices and pushing some vendors into negative gross margins.

DeepSeek slashed prices on its V4-Pro model by 75%, but for many enterprise AI vendors the cut has not translated into healthier margins — because agent-based systems are consuming tokens far faster than prices are falling [1].

The core issue, which VentureBeat’s sources describe as the “100x problem,” is that a single user-visible request routed through an agentic workflow can cost orders of magnitude more to serve than the same request handled by a simple chatbot or retrieval-augmented generation (RAG) system [1]. RAG is a technique that supplements a model’s response with text retrieved from an external knowledge base.

How token amplification works

In a standard single-turn chatbot, one user message produces roughly one model call, yielding an input-to-billed-token ratio of about 1:5 [1]. In a multi-step agent deployed across customer support, sales, finance, legal, or engineering, that ratio routinely reaches 1:700 or higher, according to the analysis [1].

The reason is structural: every loop iteration in an agent carries forward the full cumulative conversation, tool outputs, and reasoning traces, and each step appends rather than replaces [1]. A query as straightforward as “What did our top customer ask about last week?” can touch seven separately priced operations — including multiple model calls, a retrieval step, and tool execution — before returning an answer, totaling roughly 35,000 input tokens billed [1].

At frontier-model pricing, that single query costs somewhere between $0.10 and $0.40 [1]. At one million queries a month — described as “table-stakes volume for any enterprise B2B feature” — the line item reaches six figures [1].

The seat-based SaaS model under pressure

The dominant enterprise AI pricing structure has been seat-based software-as-a-service (SaaS): a flat per-user monthly fee in exchange for agent capability [1]. That model assumes a reasonably bounded cost per user, an assumption token amplification breaks [1].

A power user running 50 agent invocations a day on a $40-per-seat plan can generate more in inference costs than the plan charges, pushing vendor gross margins negative [1]. Several vendors are now privately reporting exactly that outcome on heavy users, mirroring cloud expenditure data from the Bessemer “Supernova” cohort, where the correlation between AI-agent adoption and gross margin contraction has moved from a theoretical risk to a primary profit-and-loss headwind, according to the report [1].

The dynamic creates a counterintuitive problem: the customers extracting the most value from a product are often the ones generating the highest inference costs, meaning the most engaged users can contribute the least profit [1].

Bryan Catanzaro, VP of Applied Deep Learning at Nvidia, put the pressure in stark terms: “For my team, the cost of compute is far beyond the costs of the employees.” [1]

Bloomberg separately documented a widening gap between Salesforce’s Agentforce marketing demos and the capabilities actually shipping to customers — the kind of gap, the VentureBeat analysis argues, that opens when promised functionality is technically possible but uneconomical to serve at the price a seat plan implies [1].

Technical responses and what enterprises should do

Several mitigation techniques are available, though none eliminates the underlying structural problem [1]. Cost-aware routing — using a small classifier model to direct each query to the appropriate model tier — can cut inference bills by around 60% without quality degradation, according to the analysis [1]. Prompt caching, now offered by Anthropic, OpenAI, and Google at discounts of 75 to 90% on cached prefixes, is another lever [1]. Context discipline — truncating tool outputs, pruning reasoning traces, and capping tool depth — prevents agents from running unbounded loops [1].

IBM research cited in the piece found that “organizations using orchestration-led governance report stronger productivity gains — a holistic orchestration layer is associated with six times greater productivity impact than compliance-only approaches.” [1]

The analysis recommends that enterprise leaders treat inference cost as a first-class metric tracked per feature, per tenant, and per query class; set hard cost-per-thousand-queries ceilings; treat the routing layer as core infrastructure rather than an optional optimization; audit system prompts quarterly; and negotiate volume-commit deals with model providers, since list price is described as “the worst price any enterprise will ever pay” [1].

Frontier inference unit costs are dropping roughly 3x per year and the curve is not slowing, the authors note — but amplification is outrunning those cuts, meaning a 75% price reduction does not help a company whose agents are performing 700x more tokens per query than its pricing model assumed [1].


Sources

  1. VentureBeat — DeepSeek cut prices 75%. The 100x problem remains

This article was drafted with AI from the cited sources and checked against them before publication. Spot an error? Let us know.

An analysis of 1.2 million anonymized sessions shows business operations and content creation dominate the AI chat agent's workload — not coding.

Business operations and content creation account for roughly half of all activity on Claude Cowork, Anthropic’s AI chat agent, according to an analysis of 1.2 million anonymized sessions recorded between May 11 and 31, 2026, across more than 600,000 organizations [1].

The finding matters for developers and enterprise buyers because it suggests the interface — not just the underlying model — shapes how workers actually adopt AI tools [1]. Cowork was built to bring agent-like capabilities to a familiar chat window, and the usage data shows it is being pulled toward the unglamorous coordination work that fills most office jobs [1].

What people are actually doing

Anthropic used an automated system to assign each session to one of 20 work categories [1]. “Business process and operations” came in first at 33.4 percent, covering tasks like pulling scattered status updates into a single report, building onboarding checklists, and reconciling spreadsheets [1]. “Content creation and copywriting” ranked second at 16.4 percent, spanning drafts, slide decks, posts, and proposals [1].

Anthropic describes these activities as “connective in nature,” noting that spreadsheets consolidate scattered data, slide decks translate decisions for audiences with varying backgrounds, and onboarding checklists pass along institutional knowledge to new hires [1].

The remaining categories are considerably smaller. Software development accounts for 8.7 percent, DevOps and infrastructure for 7 percent, research for 6.4 percent, data analysis for 5.8 percent, document processing for 4.1 percent, and sales operations for 4 percent [1]. Personal assistance sits at 3.8 percent, education at 2.4 percent, and meeting analysis at 1.8 percent [1].

Why coding ranks low

The relatively small share of coding sessions — 8.7 percent — is intentional, according to Anthropic [1]. Developers still use Claude Code, a separate product, to build, debug, and ship software; in Cowork, they handle the communication and coordination tasks that surround that work instead [1].

Anthropic points to a parallel pattern that emerged after Claude Code launched in 2025, when a notable number of non-technical users showed up in the terminal to organize folders, remove duplicate files, and write spreadsheet formulas — tasks that had nothing to do with writing code [1]. Cowork was designed to serve those users without requiring them to touch a command line [1].

Limits of the data

Anthropic acknowledges several constraints in the analysis [1]. The sample captures only a fixed maximum number of sessions per hour, meaning peak usage periods are underrepresented [1]. The figures are percentages rather than absolute session volumes, and roughly five percent of sessions involve personal use such as hobbies or casual chat, so the dataset does not reflect purely professional activity [1].

Anthropic also notes that its “Business Ops” category classifies activities rather than job functions — there are no separate buckets for marketing, finance, or human resources — which likely inflates that category’s share, the company said [1].

Cowork has recently expanded beyond the desktop application to the web and smartphones, and Anthropic had previously given the agent the ability to directly operate Mac and Windows desktops [1]. The company said it will continue updating the usage analysis [1].


Sources

  1. The Decoder — Claude Cowork's biggest use case is the mundane office work nobody wants to own, Anthropic says

This article was drafted with AI from the cited sources and checked against them before publication. Spot an error? Let us know.

The desktop-only AI agent now runs cross-device and in the cloud, with usage data showing knowledge workers — not developers — are its biggest audience.

Anthropic on Tuesday expanded Claude Cowork — its AI agent for general knowledge work — to iOS, Android, and the web for the first time, while also publishing usage data showing that software development accounts for less than 9% of what people actually do with the tool. [1,2]

Until now, Cowork was only accessible through the Claude desktop app for macOS and Windows. [1] The rollout begins in beta with Max subscribers and will extend to other plans “in the coming weeks,” the company said. [1,2]

What’s new in the expansion

Three concrete capabilities come with the mobile and web launch. Sessions now sync across devices, so a task started on a laptop can be checked and retrieved from a phone or browser. [2] More significantly, Cowork can now run tasks in the background with no device online at all — users can schedule work for a specific time and Claude will execute it autonomously. [2,3] Anthropic’s announcement offers a concrete example: “Set Monday’s client prep for 6 am: Claude works through the email threads, transcripts, and recent news, builds the briefing doc, and leaves the follow-up email drafted but unsent. Review it over coffee.” [2,3]

Third, when Claude reaches a decision point that requires human judgment, it surfaces the question to the user’s phone. [2,4] “Nothing ships until you’ve reviewed and approved it,” Anthropic states. [2]

The desktop app remains the most fully featured surface, retaining local file access, browser control, and Computer Use — the capability that lets Claude click, type, and navigate directly on screen. [1,4] The web version opens Cowork to users in enterprise environments where IT departments restrict software installation. [2,4]

On web and desktop, the Chat and Cowork interfaces now share a single home screen, with projects and artifacts persisting across both modes. [1,4] Anthropic is also extending its doubled Cowork usage limits through August 5. [1,2]

Who’s actually using it — and for what

Alongside the launch, Anthropic published usage data drawn from 1.2 million anonymized Cowork sessions sampled between May 11 and May 31, across more than 600,000 organizations. [2,3]

Business process and operations — tasks like pulling scattered updates into a single report, building onboarding checklists, and reconciling spreadsheets — was the single largest category at 33.4% of sampled sessions. [2,3] Content creation and copywriting came second at 16.4%, covering drafts, slide decks, social posts, and proposals. [2,3] Together, those two categories account for roughly half of all Cowork usage. [2,4]

Software development, by contrast, accounted for just 8.7% of sessions. [2,3] DevOps and infrastructure followed at 7%, research and intelligence at 6.4%, and data analysis at 5.8%. [2]

Anthropic describes the dominant use cases as “the work around the work” — tasks that span nearly every role but rarely appear in anyone’s core job description. [2,3] “While coding is still — understandably — one of the uses of AI that gets the most attention, the use of AI for everyday business work is on the rise,” the company said. [3]

Anthropic is transparent about the data’s limits: the sample is rate-capped rather than proportional to total traffic, meaning peak-hour usage is somewhat underrepresented, and roughly 5% of sessions involved personal rather than work use. [2] The company also noted that its labeling pipeline changed around May 11, which is why the analysis window starts on that date. [2]

Broader context: security concerns and competitive moves

The expansion arrives against a backdrop of unresolved security questions. On July 1, security firm Armadin — led by Mandiant founder Kevin Mandia — published research describing what it called a full sandbox escape in Claude Cowork on Windows, involving DLL sideloading against the Claude desktop executable to gain trusted access to Cowork’s virtual machine service, according to VentureBeat’s reporting. [2] Anthropic responded that the vulnerability did not qualify as a security issue because exploiting it requires an attacker to already have local code execution on the host machine. [2] The web and mobile versions run tasks server-side rather than in a local virtual machine, which removes that specific attack surface but introduces different questions around data handling for scheduled background tasks. [2]

Cowork launched in January alongside an explicit warning from Anthropic about prompt injection risks, with the company noting at the time: “These risks aren’t new with Cowork, but it might be the first time you’re using a more advanced tool that moves beyond a simple conversation.” [2]

On the competitive side, TechCrunch noted that OpenAI has made a similar push with Codex, which began as a software development tool but is increasingly being used by non-developers for reports, spreadsheets, and research. [3] The Decoder observed that Anthropic appears headed toward merging Chat and Cowork into a single product, much as OpenAI plans to do with Codex and ChatGPT. [4]

Cowork runs on Claude Sonnet 5, a model Anthropic launched last week at introductory pricing of $2 per million input tokens through August 31, before rising to $3 per million input tokens. [2]


Sources

  1. The Verge — Anthropic is launching Claude Cowork on mobile and web
  2. VentureBeat — Anthropic brings Claude Cowork to mobile and web as usage data shows most users aren’t coding
  3. TechCrunch — Claude Cowork expands to mobile and web
  4. The Decoder — Anthropic's Claude Cowork AI agent is now available on mobile and web

This article was drafted with AI from the cited sources and checked against them before publication. Spot an error? Let us know.

An internal memo reveals Microsoft is stripping back Copilot to focus on "real work," combining consumer and enterprise apps into one and adding background-task agents.

Microsoft plans to release a heavily rebuilt version of Copilot in August that merges its consumer and enterprise apps into a single product and introduces new AI agents called “AutoPilot,” according to an internal memo seen by The Information [1].

The move matters for developers and business buyers because it signals Microsoft is betting that bundled, outcome-focused tooling — rather than a standalone chatbot — is what justifies the cost of enterprise AI [1].

What’s changing inside Copilot

Executive Vice President Jacob Andreou authored the memo, writing that the team “stripped out what wasn’t working,” naming Copilot Podcasts and Copilot Labs as features being cut [1]. Andreou framed the new direction around practical results: Copilot should focus on “real work” instead of chasing intelligence “for intelligence’s sake,” and must be “optimized for outcomes,” he wrote [1]. In pointed language, he added that the app has to “earn the right to exist” [1].

The rebuilt app will include AI coding tools alongside the AutoPilot agents, which are designed to handle background tasks such as scheduling and email summaries [1]. Microsoft will charge extra for those added features, according to the memo [1].

A broader push beyond the chatbot

The Copilot overhaul is not Microsoft’s only move this week. The company also announced a new unit in which Microsoft engineers will work directly inside client departments to embed AI into business workflows, The Decoder reported [1]. That step is, as The Decoder noted, an acknowledgment that a chatbot alone delivers limited — or hard-to-measure — value [1].

Microsoft is not alone in pursuing what The Decoder describes as an AI “super app.” Anthropic, with Claude Code, and OpenAI, with Codex, are working on comparable all-in-one products, the outlet reported [1]. All three companies are still under pressure to demonstrate returns on their substantial AI infrastructure spending [1].

The revamped Copilot is reportedly targeted for an August release [1].


Sources

  1. The Decoder — Microsoft follows Anthropic and OpenAI into the AI super app race with overhauled Copilot and AutoPilot agents

This article was drafted with AI from the cited sources and checked against them before publication. Spot an error? Let us know.

Thariq Shihipar argues that with Fable 5, output quality is now bottlenecked by the user's ability to surface their own blind spots — not by the model.

An Anthropic developer has published a detailed prompting guide for Claude’s latest model, Fable 5, arguing that the biggest barrier to good output is no longer the model itself but the user’s unexamined assumptions [1].

Thariq Shihipar, writing about his own agentic coding workflow, says Fable 5 is “the first model where output quality is limited by the user’s ability to clarify their ‘unknowns'” [1]. For developers building with AI agents, that reframes the core problem: less about coaxing the model and more about interrogating your own thinking before you write a single line of a prompt [1].

A Four-Category Framework for What You Don’t Know

Shihipar organizes the problem using four categories [1]. “Known Knowns” are facts already in the prompt. “Known Unknowns” are gaps you’re aware of. “Unknown Knowns” are things so obvious you’d never write them down. The most dangerous category, he argues, is “Unknown Unknowns” — things you haven’t considered at all [1].

Getting the prompt specificity wrong in either direction causes failures, according to Shihipar [1]. Too much detail and Fable 5 rigidly follows instructions even when a course correction would be better; too little and the model falls back on generic industry defaults that may not fit the task [1]. “When you don’t account for your unknowns you fail both ways,” he writes [1].

Pre-Implementation Techniques

Shihipar describes several techniques for the phase before any code is written [1]. The first is a “blindspot pass,” in which the developer explicitly asks Claude to identify unknown unknowns — particularly useful when working in an unfamiliar part of a codebase [1]. An example prompt he offers: “I’m working on adding a new auth provider but I know nothing about the auth modules in this codebase. Can you do a blindspot pass to help me figure out my relevant unknown unknowns and help me prompt you better.” [1]

For domains heavy in “Unknown Knowns” — he cites visual design as an example — Shihipar recommends generating several radically different directions as HTML artifacts to react to, rather than jumping straight to implementation [1]. He also describes “structured interviews,” where Claude asks the developer question by question about ambiguities, prioritizing questions whose answers would most affect the architecture [1].

Source code, even in a different programming language, is his preferred reference material [1]. He notes that Claude Design reads a website’s underlying code rather than a screenshot [1]. Before work begins, he has Claude produce an implementation plan focused on the parts most likely to change — data models, type interfaces, and user-facing elements — leaving mechanical refactoring for last [1].

During and After Implementation

Unknowns don’t stop surfacing once coding starts, Shihipar warns [1]. He asks Claude Code to maintain a temporary “implementation-notes.md” file tracking decisions as they’re made, so the team can learn from each attempt [1]. When unexpected edge cases appear, his rule is for Claude to pick the conservative option, log the deviation, and keep moving [1].

After implementation, he uses two additional techniques [1]. “Pitches and explainers” are summary documents that bundle the prototype, specs, and implementation notes for stakeholders [1]. “Quizzes” have Claude generate an HTML report on the changes made, followed by a quiz — Shihipar says he does not merge code until he passes the quiz without errors [1].

He illustrates the full workflow with the launch video for Fable, which he edited entirely using Claude Code despite having no prior video editing experience [1]. When the result looked flat in color grading, he did not simply ask Claude to generate variations; he first had Claude teach him what good color grading looks like, in order to surface his own unknowns before evaluating options [1].

“Every explainer, brainstorm, interview, prototype, and reference is a cheap way to find out what you didn’t know before it gets expensive to fix,” Shihipar writes [1]. He has also published a visual version of the framework on a dedicated website [1].


Sources

  1. The Decoder — Anthropic developer shares prompting tips for Fable 5 that focus on finding your own blind spots first

This article was drafted with AI from the cited sources and checked against them before publication. Spot an error? Let us know.

A new framework from Alibaba researchers decomposes complex tasks iteratively so AI agents fetch only the tools they need, slashing context costs from ~884,000 tokens to ~1,160 per query.

Alibaba researchers have published SkillWeaver, a framework that routes AI agents to the right tools through iterative task decomposition rather than loading an entire tool library into context, reducing token consumption by over 99% compared to brute-force approaches [1].

For developers building enterprise agents that orchestrate dozens or hundreds of tools, the practical stakes are direct: fewer tokens mean lower API costs and faster responses, and the accuracy gains address a failure mode that currently makes large tool libraries nearly unusable [1].

The problem with loading every tool

Modern large language model (LLM) agents are increasingly integrated with massive tool ecosystems — including systems built on the Model Context Protocol (MCP), a standard for connecting agents to external services [1]. When an agent needs to handle a multi-step request like “Download the dataset, transform it, and create visual reports,” exposing the full tool library to the LLM to find the right tools is, as the researchers describe it, “highly inefficient” and “quickly overwhelms context limits” [1].

Most existing frameworks treat tool selection as a one-shot, single-skill problem, which breaks down when real business queries require sequencing multiple tools in a compatible chain [1].

How SkillWeaver works

SkillWeaver structures tool selection across three stages: Decompose, Retrieve, and Compose [1]. First, an LLM breaks a complex user query into atomic sub-tasks, each requiring exactly one tool [1]. An embedding model then searches the tool library to produce a shortlist of candidates for each sub-task [1]. Finally, a planner checks inter-tool compatibility — whether the output of one tool feeds cleanly into the next — and assembles the result as a Directed Acyclic Graph (DAG), allowing independent steps to run in parallel [1].

The central innovation is a feedback mechanism the team calls Skill-Aware Decomposition (SAD) [1]. Rather than decomposing a task once and searching, SAD has the LLM draft an initial plan, runs a preliminary search to surface loosely matching tools, and feeds those results back to the LLM so it can rewrite its decomposition using the actual vocabulary of the available tools [1]. The authors shared the prompt templates in their paper, and the technique can be reproduced with standard libraries such as LangChain, LlamaIndex, or plain Python [1].

Benchmark results

The team evaluated SkillWeaver on CompSkillBench, a custom benchmark of 300 multi-step queries run against a library of 2,209 real-world skills drawn from the public MCP ecosystem across 24 categories including cloud infrastructure, finance, and databases [1].

Without SAD, a 7-billion-parameter Qwen2.5-7B-Instruct model correctly predicted the number of required steps only 51.0% of the time; activating SAD pushed that to 67.7%, and using the larger Qwen-Max model brought it to 92% [1]. On hard tasks requiring four to five tools, SAD improved accuracy by 50% [1].

One counterintuitive finding: a 14-billion-parameter model performed worse than the 7B model in the unguided setup because it tended to over-decompose tasks into unnecessary micro-steps [1]. SAD corrected this by anchoring the larger model to the actual tool vocabulary, suggesting that tool alignment can matter more than raw model size [1].

The token savings were stark. The brute-force baseline — feeding all tool names into a large model’s prompt — consumed an estimated 884,000 tokens per query and still only retrieved the correct tool category 21.1% of the time [1]. SkillWeaver’s targeted approach brought that figure down to roughly 1,160 tokens per query, a 99.9% reduction, while substantially outperforming on accuracy [1]. A ReAct-style agent baseline achieved 0% decomposition accuracy, collapsing multi-step plans into isolated actions [1].

On retrieval infrastructure, the researchers used the open-source all-MiniLM-L6-v2 embedding model with a FAISS index [1]. Indexing all 2,209 skills took 15 seconds, and retrieval adds under 15 milliseconds of latency per query [1]. Swapping in a stronger encoder (BGE-base-en-v1.5) improved accuracy without any fine-tuning [1].

Limitations and what’s not included

The framework currently has no error-recovery mechanism [1]. If a tool call fails mid-chain, the entire execution breaks, and the authors acknowledge that practitioners will need to build their own retry and fallback logic on top of the compose stage [1].

The source code has not yet been released, though the researchers note the system was built entirely from off-the-shelf components and the SAD prompt templates are available in the paper [1].


Sources

  1. VentureBeat — New Alibaba AI framework skips loading every tool, cutting agent token use 99%

This article was drafted with AI from the cited sources and checked against them before publication. Spot an error? Let us know.

The open-source AI agent framework, which went viral earlier this year, is now available as a mobile app on both major platforms.

OpenClaw, the free, open-source artificial intelligence agent framework that gained widespread attention earlier this year, is now available as an app on both iOS and Android, the project announced on X on Tuesday. [1]

On both platforms, users can pair their phones with the OpenClaw Gateway, a routing layer that connects user requests to AI agents and the tools and skills those agents use to complete tasks. [1]

OpenClaw users have applied the framework to a range of tasks including coding and meal planning, though some have reported unsatisfactory results. [1]

The project first attracted broad attention around the launch of MoltBook, a social media site described as being populated entirely by AI agents. [1] That spectacle was later found by researchers to have been partially the work of humans impersonating agents — a move that functioned as marketing for OpenClaw, at some cost to its credibility. [1]

In February, OpenClaw creator Peter Steinberger announced he had joined OpenAI. [1]


Sources

  1. TechCrunch — OpenClaw is finally available on Android and iOS

This article was drafted with AI from the cited sources and checked against them before publication. Spot an error? Let us know.

A new macOS feature lets users demonstrate a task once and have the Codex AI agent execute it autonomously on demand.

OpenAI has added a feature called Record & Replay to its Codex macOS app, allowing users to walk the AI agent through a workflow a single time and have it convert that demonstration into a reusable “skill” it can run independently afterward. [1]

The practical use case cited in the release is uploading a YouTube video complete with metadata, a thumbnail, and subtitles — a multi-step process Codex can repeat on its own once it has been shown how. [1]

Record & Replay depends on Computer Use, a capability that lets the agent interact with a computer’s graphical interface directly. [1] Computer Use must be enabled for the feature to work, and it is not yet available in the European Union, the United Kingdom, or Switzerland — though Computer Use itself became available in the EU on June 16. [1]

Version 26.616 of the app also ships bulk actions for the Automations history view and the ability to hand off task threads between a local machine and a remote host, letting work continue across connected devices. [1]

Codex is positioned by OpenAI as an AI agent for coding and white-collar workflow automation. [1] The app is free to download, but a paid ChatGPT subscription is required for substantive use. [1]


Sources

  1. The Decoder — OpenAI's Codex can now watch you work once and repeat the task forever

This article was drafted with AI from the cited sources and checked against them before publication. Spot an error? Let us know.