Overview of Claude Co-Work can now Control Your Computer
Host Jaeden (Jaden) Schaefer walks through Anthropic’s recent Claude Co-Work update that allows Claude to literally control your computer (mouse, keyboard, screen) and a companion feature called Dispatch for remote control from your phone. He explains how it works, what he tested, practical use cases, setup tips, limitations, and who should try it—plus a short pitch for his AI platform, AIBox.ai.
Key updates covered
- Anthropic released a feature enabling Claude to control your computer (mouse/keyboard/screen) via the Claude Desktop app (initially macOS; Windows support rolled out shortly after).
- Dispatch: a companion feature that lets you monitor and control that remote Claude session from your phone.
- You can save repeatable workflows as “skills” so Claude can rerun them without re-teaching each step.
- Host tested the new features extensively and reports real-world workflows and limitations.
What Claude Co-Work and Dispatch do
- Claude Co-Work (desktop app): gives the model visibility of your screen and the ability to click/type so it can interact with native and web apps.
- Dispatch: syncs the Claude session to your phone so you can start/monitor/steer automations remotely.
- Skill saving: after refining a workflow, Claude generates a skill file that documents steps and required permissions for future runs.
Practical examples the host tried
- Automating multi-step workflows across sites (logging into sites, extracting outputs).
- Integrating with AIBox.ai: instructing Claude to call a specific “box” (automation/tool) and use its output as part of a larger workflow.
- Editing videos in CapCut: opening the app, importing files and performing edits via mouse/keyboard automation.
- Content generation: having Claude visit specific tools/pages to generate and aggregate copy, documents, or research.
Performance & limitations
- Speed: intentionally slow and cautious (described as “honey”-slow). Reliable but not fast—may feel tedious when watching step-by-step.
- Permission prompts: Claude often needs repeated permissions (clipboard, file access, browser control). Once captured in a skill, these prompts can be handled up front.
- Setup: requires the Claude desktop app (not browser-only). Some configuration and trial-and-error are needed to build robust workflows.
- Best on a separate machine: host recommends running it on an extra laptop/desktop (Mac was better initially; Windows support added).
- Not a replacement for developer tools: developers may prefer programmatic tools like Claude Code or OpenAgent/OpenClaw-style setups that can be faster/less GUI-bound.
Tips and best practices
- Use a dedicated machine (old laptop or Mac mini) to run agents so your main workstation isn’t slowed down.
- During initial runs, let Claude build the skill file — it documents required permissions and fallback approaches.
- Test steps incrementally and accept permission prompts early so later runs are smoother.
- Use AIBox.ai (host’s platform) or similar services to chain model calls; instruct Claude to call a tool/link rather than re-prompting for complex generation in-line.
- Save stable workflows as skills and name them clearly so you can "Run this skill" later.
- Expect upfront time investment — a complex 2-hour-per-day task might take many hours to automate initially but will save time long-term.
Comparison to alternatives
- OpenAI agents / other agent tools: Anthropic’s Claude Co-Work is presented as more polished for non-developers and more capable in GUI automation than many existing agent systems. Developers still may prefer code-first tools (Claude Code, OpenClaw/open-source agents) for performance or customization.
Notable quotes / host impressions
- Felix Riesenberg (Anthropic): “It’s slow, but giving Claude any mouse and keys is so exciting to me.”
- Host summary: “Feels kind of like honey”—it works and is highly useful but currently cautious/slow.
Recommendations / final takeaway
- Who should try it: white-collar professionals, non-developers, and teams with repetitive GUI-heavy tasks (newsletters, content assembly, research, editing).
- Expect an upfront setup cost (time and permissions) but significant recurring time savings once skills are refined.
- Use a dedicated machine for reliability; save skills; and integrate with tooling (e.g., AIBox.ai) to make workflows modular.
- Developers who prefer programmatic control may opt for Claude Code or other agent frameworks instead.
Resources mentioned
- Claude Co-Work and Dispatch: available via the Claude Desktop app (macOS and Windows).
- AIBox.ai: host’s platform for access to 70+ models and auto-chained automations (subscription mentioned: $8.99/month).
- Original social shoutouts referenced: Thomas Frank tweet about long automation builds and Felix Riesenberg’s Anthropic post.
If you want a condensed checklist to get started (download app, enable Co-Work & Dispatch, give permissions, build & save a skill, run on a separate machine), I can include that as a quick-start section.
