Everyone Needs to Calm Down

Everyone lost their minds over Dispatch this week. "Anthropic shipped their own OpenClaw!" "The future of personal AI!" "I can control Claude from my phone!"

I tested it. Misconfigured it on first start, connection reliability needs work, and your laptop must stay awake. It's a research preview, not a product.

But here's what's actually interesting: three companies shipped local desktop agents within the same week. Perplexity first (everyone missed this), then Anthropic, then Meta's Manus. A product category formed in real time.

The Four That Qualify Right Now

I stepped back and applied a simple filter: autonomously execute multi-step tasks, access your local files, and connect to services you actually use. Four products qualify.

Cowork + Dispatch (Anthropic)

Real VM sandbox, default-deny networking, explicit folder grants. The only commercial product with actual local isolation.

But the agent can't run in the background. Close your laptop, tasks stop. That's a fundamental limitation that makes "delegate from your phone" feel more like a demo than a workflow.

Strengths: Security model, reasoning quality, structured sandbox. Weaknesses: No background execution, laptop-dependent, research preview stability.

Manus (Meta)

Strong general-purpose agent, $2B acquisition backing. But no sandbox. Every CLI command goes through an "Allow Once / Always Allow" prompt.

One "Always Allow" click and the guardrail is gone. That's not a security model – it's a liability waiver disguised as a permission dialog.

Strengths: General-purpose capability, Meta's resources and model ecosystem. Weaknesses: No isolation, permission model is security theater, provider lock-in.

Perplexity Computer

Background daemon (Mac-only), multi-model routing, cron-like recurring tasks, multi-day persistence. This is the feature set everyone wanted from Dispatch but didn't get.

Local sandbox situation is undocumented. That's either "we haven't built it yet" or "we built it but it's not ready for scrutiny." Neither is reassuring.

Strengths: Background execution, recurring tasks, multi-model routing, persistence. Weaknesses: Mac-only, undocumented security model, early-stage.

DIY (OpenClaw et al.)

The only option that can run fully local with local models. 13,700+ skills. Full control over everything.

But nothing is configured by default. No security, no sandbox, known CVEs, supply-chain attacks. If you can't run a command line, stay away.

Strengths: Full local execution, model choice freedom, maximum optionality, 13,700+ skills. Weaknesses: No default security, known vulnerabilities, requires deep technical skill.

The Comparison Matrix

Capability	Cowork + Dispatch	Manus	Perplexity Computer	DIY (OpenClaw)
Local file access	✅ Folder grants	✅ Full	✅ Full	✅ Full
Sandbox isolation	✅ VM-based	❌ None	❓ Undocumented	❌ Manual setup
Background execution	❌ Laptop must stay awake	✅	✅ Daemon	✅ Self-hosted
Multi-model routing	❌ Claude only	❌ Meta models	✅ Multiple	✅ Any model
Recurring tasks	❌	❌	✅ Cron-like	✅ Full control
Fully local option	❌	❌	❌	✅
Works out of the box	⚠️ Research preview	✅	✅	❌
Security posture	🟢 Strong	🔴 Weak	🟡 Unknown	🔴 Your problem

Two Paths, Not Four Products

The real choice isn't between four products. It's between two fundamental paths.

Path 1: DIY

Maximum optionality, maximum responsibility. You choose your model (local or cloud), your security posture, your privacy model, your integrations. The ceiling is highest. The floor is lowest. Every misconfiguration is on you.

This path makes sense if you:

• Need data to never leave your network
• Want to use specific local models (Qwen, Nemotron, etc.)
• Have the technical skills to configure and maintain the stack
• Value control over convenience

Path 2: Commercial

Opinionated, provider-locked, works out of the box. The provider chose the model (theirs), the security model (their sandbox or lack thereof), and the privacy posture (your data transits their cloud for reasoning). You get a working product fast, but you're constrained by their decisions.

This path makes sense if you:

• Need something working today
• Trust the provider's security model
• Don't need full local execution
• Value time-to-value over control

What This Means for Businesses

For our clients at Till Freitag, the advice is nuanced:

If you're exploring agents for the first time: Start with Cowork + Dispatch. The security model is real, and the friction will teach you what agents can and can't do yet.

If you need background automation: Watch Perplexity Computer closely. The daemon architecture is what production agent workflows actually need.

If data sovereignty is non-negotiable: DIY is your only option. But invest in the security layer first – not after you've been running unpatched for three months.

If you're evaluating for a team: Wait. None of these are enterprise-ready. The permission models, audit trails, and multi-user scenarios are all unsolved.

The Bottom Line

A product category formed in one week. No product leads on all four axes: security, background execution, model flexibility, and ease of use.

The category is weeks old. Pick the tradeoff you can live with, not the one that sounds best on a LinkedIn post.

→ Read our Anthropic Dispatch analysis → Privacy Router: How to route sensitive data locally → OpenClaw alternatives compared