I tried Claude’s new Cowork feature — and it ran my laptop from my phone

Anthropic released Dispatch and Cowork (computer use) for Claude, enabling the model to control a user's desktop (move cursor, open apps, type, navigate) and run tasks remotely from a phone; the features are in research preview and require Claude Pro/Marx plus the latest desktop and mobile apps. The capability can materially speed productivity by executing background tasks, but it requires broad permissions (file access, mouse/keyboard control, app access) and carries accuracy and data-privacy risks, so outputs must be reviewed.

Analysis

Anthropic’s “Dispatch” and desktop control features accelerate a shift from real‑time conversational assistance to unattended task automation — a structural change that should increase spending on endpoint security, identity, and audit tooling within 6–24 months. Enterprises that allow an LLM to control a user’s desktop will demand stronger telemetry, immutable audit trails, and runtime policy enforcement; that drives incremental ARR for vendors who can certify secure, auditable execution environments.

Second‑order winners are RPA and orchestration vendors that can embed LLM‑driven flows into existing automation pipelines: instead of building bespoke connectors, customers will want orchestrators that gate and log desktop actions. Conversely, consumer LLMs and unsupervised desktop control raise immediate liability and compliance risk — a single accidental data exfiltration or mis‑typed financial instruction could trigger class actions or regulatory investigations that slow enterprise adoption for 3–12 months.

Catalysts to watch: enterprise pilot wins and SOC/IR vendor partnerships (near term, weeks–months); first high‑profile security incident or regulator inquiry (fast negative catalyst, days–weeks); official product integrations from hyperscalers or on‑prem deployment options (3–12 months positive). The contrarian angle: adoption will likely be bumpy and selective — knowledge workers in finance/healthcare will block it initially, meaning near‑term revenue flow is concentrated in security/observability sellers rather than LLM vendors themselves.

Key Decisions for Investors

• Overweight CRWD (CrowdStrike) — buy and hold 6–12 months. Rationale: endpoint control + audit demand should lift ARR and gross retention. Target +20–30% if adoption accelerates; downside 20–30% if no incremental spend materializes or multiples compress after a macro pullback.
• Buy PATH (UiPath) on any <10% pullback, 3–9 month horizon. Rationale: RPA vendors become the path of least resistance to fold dispatched desktop actions into enterprise workflows. Reward: ~30% upside if enterprise pilots convert; Risk: ~40% downside if LLM vendors internalize orchestration.
• Directional GPU / infra exposure — long NVDA via 12–24 month call spread (e.g., buy calls, sell higher strike) sized for 6–10% portfolio exposure. Rationale: on‑prem / hybrid LLM inference demand rises if enterprises distrust cloud desktop control. Reward: large upside from continued GPU tightness; Risk: cloud optimization or custom accelerators reduce demand growth.
• Hedge idea for GOOGL holders — add CRWD or OKTA exposure rather than selling GOOGL outright. If you prefer explicit hedge, implement a market‑neutral pair: long CRWD / short GOOGL (size 1:0.25) for 6–12 months to capture security premium while limiting big‑cap directional risk. Potential reward: outperformance if security multiple expands; risk: GOOGL outperforms broadly and compresses pair.