Aller au contenu principal
Main translucide lumineuse traversant un écran pour contrôler un curseur, métaphore de Claude Computer Use

Claude in March 2026: Computer Use, Dispatch and all Anthropic updates

Back to blog
Artificial Intelligence
Nicolas
11 min read
Main translucide lumineuse traversant un écran pour contrôler un curseur, métaphore de Claude Computer Use

On March 23, 2026, a 90-second video went viral: a user controls their Mac from their phone, delegates the task of compiling a report to Claude Dispatch, then watches the computer act on its own.

25 million views in nine hours.

This is no accident: it’s the culmination of three months of releases that made Q1 2026 the most prolific quarter in Anthropic’s history.

New models, a stabilized Computer Use, Dispatch, Remote Control, Agent Teams, MCP Apps: each piece slots into a clear vision of turning Claude into an autonomous workstation.

Key takeaways:

  • Claude Sonnet 4.6 hits 72.5% on OSWorld, on par with a human operator (70-75%)
  • Dispatch controls your computer remotely from your phone, with no cloud server in between
  • The approach is local-first: your data never leaves your machine, but it must stay on
  • Anthropic reaches 11.3 million daily active users (+183% since January 2026)
  • The convergence of Claude Code + Cowork + Dispatch marks the first coherent AI work OS on the market

January-March 2026: the complete timeline of Claude updates

Here are all the major releases of the quarter, in chronological order:

Date Feature Availability
January 2026 Interactive MCP Apps (Slack, Figma, Asana, Canva, Box) Claude Pro/Max
Feb. 5, 2026 Claude Opus 4.6 (1M token context in beta, advanced coding) Pro/Max/API
Feb. 17, 2026 Claude Sonnet 4.6 (Opus 4.5 performance at Sonnet pricing, $3/M tokens) Pro/Max/API
Feb. 20, 2026 Claude Code Security (vulnerability scanning, targeted patches) Claude Code
Feb. 25, 2026 Claude Remote Control (early access CLI for developers) Claude Code
Mar. 12, 2026 Claude Partner Network ($100M enterprise partner investment) Enterprise
March 2026 Claude Dispatch (research preview, Max then Pro) macOS only
March 2026 Agent Teams in Claude Code (parallel multi-agents with direct communication) Claude Code

This timeline is not a collection of unrelated features: Remote Control in February paves the way for Dispatch in March, Opus 4.6 makes sustained agentic behavior possible, and Sonnet 4.6 brings those capabilities to the entry-level price point.

Each release this quarter prepares the next.

The timeline is not a launch schedule: it’s a construction plan.

Computer Use: how Claude sees and controls your Mac

Computer Use is the technical engine behind Dispatch, Cowork and Remote Control: it’s Claude’s ability to perceive the screen and interact with interfaces.

It operates on a continuous vision-action loop: Claude captures a screenshot, analyzes the interface state, determines the action to take (click, type, scroll, drag), executes it, then starts again from a fresh screenshot.

Pixel-level precision was the main training challenge: Claude must calculate exact X/Y coordinates to click in the right spot, a skill that early large language models couldn’t reliably master.

The benchmark results speak for themselves:

  • OSWorld (Ubuntu/Windows/macOS desktop tasks): 14.9% at public beta launch (Oct. 2024) → 28% with Claude 3.7 Sonnet → 72.5% with Sonnet 4.6, with a human baseline of 70-75%
  • WebVoyager (multi-site web automation): 93.9% success rate for the Magnitude/Claude implementation, vs. 87.0% for OpenAI Operator, 89.1% for browser-use, and 85.9% for Skyvern

Computer Use is available exclusively for Claude Pro and Max subscribers, on macOS.

For a solid grasp of the security implications, our in-depth analysis of Computer Use is still a worthwhile read.

The real limitations are worth naming clearly:

  • Latency is noticeable (each vision-action loop takes several seconds, even longer via Dispatch)
  • Non-standard interfaces or complex visual overlays trigger click errors
  • The system remains vulnerable to prompt injection attacks: a malicious email opened during a session can trigger unintended actions
  • Operations requiring system administrator rights are out of reach

Anthropic documents these limitations openly and recommends not exposing sensitive data while the feature is still maturing.

Silhouette géométrique de chef d'orchestre dirigeant des flux de communication colorés, métaphore de Claude Dispatch

Remote Control and Dispatch: two ways to operate remotely

Anthropic launched two implementations of the same core idea, aimed at two different audiences.

Claude Remote Control (for developers)

Available in early access on February 25, 2026, Remote Control is activated via the command line: claude remote-control.

The command creates a persistent session on the local machine, accessible from Claude Code’s web, iOS and desktop interfaces.

This is the option for developers comfortable with the terminal who want to run complex workflows remotely, with fine-grained control over the permissions they grant.

Claude Dispatch (for everyone)

Dispatch is the consumer version of Remote Control: same technology, radically different interface.

Setup takes seconds by scanning a QR code in the Cowork section of Claude Desktop, with no command line or technical installation required.

Once paired, the user sends instructions from their phone and Claude executes them on the desktop computer.

Think of an autonomous colleague you hand the office keys and a task list to: they work while you’re on the road, without your data passing through any third-party server.

A few concrete cases from early Pro/Max user feedback:

Legal: triggering the compilation and formatting of every termination clause in a full case file, directly from a phone.

Marketing: pulling Amplitude KPIs, creating a Canva slide and a coordinated Slack summary, all from a single message sent from a phone.

For a full picture of what this feature covers, our detailed breakdown of Claude Dispatch walks through the setup, use cases and exact limitations.

Critical constraint: the computer must stay on and Claude Desktop must remain open.

No processing takes place on Anthropic’s servers: if your Mac goes to sleep, nothing runs.

End-of-task notifications are not yet available: you need to manually check the Dispatch conversation thread.

MCP Apps, Cowork and Agent Teams: the complete agentic toolkit

MCP Apps: Claude as your central command point

Launched in January 2026, interactive MCP Apps give Claude direct access to third-party services without leaving the interface.

Available integrations at launch: Slack, Figma, Asana, Box, Canva.

Access is bidirectional: Claude doesn’t just read data: it can edit Figma designs, send Slack messages, update Asana tasks or move Box files.

The MCP (Model Context Protocol), launched by Anthropic in 2024, now counts over 300 pre-built MCP servers, available in one click via Docker Desktop.

Cowork: the no-code interface for non-developers

Cowork is the abstraction layer that makes agentic capabilities accessible without a command line.

File access works through designated folders: the user selects the directories Claude can access, and the system never steps outside those boundaries.

Scheduled tasks enable automatic recurrences (daily email summaries, weekly reports from Google Drive), as long as the computer stays on and Claude Desktop remains open.

Cowork runs on the Claude Agent SDK, which ensures that improvements to the base model automatically benefit Cowork and Claude Code users.

Claude Code: multi-agent coordination with Agent Teams

The Agent Teams update of March 2026 introduces a multi-agent architecture within Claude Code.

A team lead agent coordinates multiple “teammate” agents working in parallel, with direct agent-to-agent communication via a shared task list.

Teammates can self-assign available tasks or receive assignments from the lead, with dependency tracking: a blocked task won’t start until its prerequisites are complete.

Two display modes: in-process (single terminal, keyboard navigation) or split pane (each agent in its own panel, via tmux or iTerm2).

The arc mirrors the iPhone’s: a basic phone in 2007, an app platform in 2008.

Claude follows the same trajectory: a chatbot in 2023, an agentic work OS in 2026.

March 2026 comparison: Claude vs. five competing approaches

System Approach Scope Price Strength Main limitation
Claude Dispatch/Cowork Local-first Full desktop $20/month (Pro) Best desktop + web benchmarks (93.9%) Computer must stay on
OpenAI Operator Cloud (VM) Web only $200/month Zero local setup, easy access 87% WebVoyager, web only
Perplexity Computer Hybrid local/cloud Desktop (Mac Mini) N/A Multi-model, local GDPR infrastructure Requires dedicated hardware
Google Mariner Cloud (Gemini) Web Gemini Ultra Native Google Workspace integration Closed beta, web-only scope
OpenClaw Open source Desktop Free Customizable, self-hostable Complex setup, no official support

Claude’s decisive edge comes from combining full desktop scope + performance (72.5% OSWorld) + accessible pricing ($20/month): no competitor checks all three boxes.

OpenAI Operator takes the opposite approach: zero local infrastructure for the user, confined to the web, and priced ten times higher.

The adoption question goes beyond features: the mass exodus from ChatGPT to Claude observed in early 2026 is partly explained by this perceived functional advantage around data control.

The local-first approach positions Claude as the only consumer solution that combines benchmark performance and data privacy without dedicated hardware.

Cinq agents IA lumineux en formation pentagonale dans l'espace cosmique, métaphore de Claude Agent Teams

Concrete impact for professionals and businesses

Use cases emerging from early Pro/Max user feedback point to three recurring patterns.

Reinforced async work: delegate from your phone while commuting, find the finished result when you get to the office.

Cross-tool synthesis automation: combine data from Google Sheets, Slack, Notion and generate a report without manual copy-pasting.

Secure code review: Claude Code Security scans entire codebases for logical vulnerabilities that rule-based tools consistently miss.

Availability remains restricted for now:

  • Computer Use, Dispatch and Cowork: Claude Pro ($20/month) and Claude Max ($100/month) subscribers, on macOS only
  • Claude Code and Remote Control: accessible via the API and CLI tool, no Pro subscription required for developers
  • Agent Teams: available in Claude Code for any user with access to the tool

For European companies subject to GDPR, the local-first approach is a direct commercial argument: data never leaves the user’s machine.

The launch of the Claude Partner Network ($100 million invested) supports this enterprise adoption with dedicated training and technical support.

The Anthropic strategy behind the Q1 2026 acceleration

The growth figures capture the scale of the shift: 11.3 million daily active users, up 183% since January 2026.

Anthropic’s valuation has reached $380 billion following a $30 billion funding round.

Claude Code alone has an annualized revenue approaching $1 billion, less than a year after its launch in May 2025.

The strategic coherence rests on a strong architectural choice: the local-first model.

Unlike OpenAI Operator, which runs tasks in the cloud, Claude runs everything locally on the user’s machine.

This philosophy ties all the pieces together: Claude Code, Cowork, Dispatch, Remote Control, Agent Teams.

Each is an additional abstraction layer on the same engine, made accessible to a broader audience.

The result: one agent, multiple interfaces, no mandatory cloud infrastructure.

If you have a Claude Pro or Max subscription on Mac, Computer Use and Dispatch are available right now in Claude Desktop, in the Cowork section.

The best starting point: begin with a simple task (summarizing files in a folder, compiling a report) before expanding the permissions you grant Claude.

FAQ

What is Claude Computer Use and who can access it?

Computer Use is Claude’s ability to see your computer screen and interact with interfaces through a vision-action loop.

It’s available for Claude Pro ($20/month) and Max ($100/month) subscribers on macOS only.

What’s the difference between Dispatch and Remote Control?

Remote Control is activated via the command line (claude remote-control) and targets developers comfortable with the terminal.

Dispatch is the consumer version: QR code setup, mobile interface, no terminal required, with the same local execution capabilities.

Does my computer need to stay on for Dispatch to work?

Yes, this is a fundamental constraint of the local-first model: no processing takes place on Anthropic’s servers.

If your Mac goes to sleep or Claude Desktop is closed, tasks won’t run.

Is Claude Dispatch available on Windows or Linux?

As of March 2026, Dispatch and Computer Use are restricted to macOS.

Anthropic has not shared a timeline for other platforms.

What are the real Computer Use benchmarks?

On OSWorld (desktop tasks), Claude Sonnet 4.6 reaches 72.5%, compared to 70-75% for human operators.

On WebVoyager (web automation), the Magnitude/Claude implementation reaches 93.9%, vs. 87% for OpenAI Operator.

Which apps are available through MCP Apps in March 2026?

Available integrations at launch include Slack, Figma, Asana, Box and Canva.

Access is bidirectional: Claude can read and modify content across these services directly from its interface.

Is Claude Dispatch secure for professional data?

The local-first model ensures your data never passes through Anthropic’s servers.

The main risk is a compromised paired phone: Anthropic recommends carefully managing permissions and knowing how to revoke access quickly in case of an incident.

What is Claude Cowork and how does it differ from Claude Code?

Cowork is the no-code version of Claude’s agentic capabilities, accessible via a graphical interface with folder selection.

Claude Code is the CLI tool for developers: Cowork targets non-technical professionals who want to automate tasks without writing code.

What is Anthropic’s strategy behind this wave of releases?

Anthropic is building a coherent agentic work platform where each release (Opus 4.6, Sonnet 4.6, Remote Control, Dispatch, Agent Teams) builds on the one before.

The goal is to make Claude an autonomous workstation, not a chatbot with added features.

How do I get started with Computer Use and Dispatch with a Pro or Max subscription?

Update Claude Desktop (macOS) to the latest version, then open the Cowork section.

For Dispatch, scan the QR code from the Claude mobile app: for Computer Use, grant access permissions for the apps you want from Cowork preferences, starting with the least sensitive ones.

Related Articles

Ready to scale your business?

Anthem Creation supports you in your AI transformation

Disponibilité : 1 nouveau projet pour Avril/Mai
Book a discovery call
Une question ?
✉️

Encore quelques questions ?

Laissez-moi votre email pour qu'on puisse continuer cette conversation. Promis, je garde ça précieusement (et je ne vous bombarderai pas de newsletters).

  • 💬 Accès illimité au chatbot
  • 🚀 Des réponses plus poussées
  • 🔐 Vos données restent entre nous
Cette réponse vous a-t-elle aidé ? Merci !