Every cloud agent tool you have tried this year was blind to your actual files and locked out of your live browser sessions.

The workaround era for desktop AI just ended

Knowledge workers have spent two years copy-pasting content into web chat windows because cloud agents cannot touch local files or authenticated browser tabs. That gap kills the automation before it starts.

A 300-agent swarm now runs on your own machine

Moonshot AI Launches Kimi Work, a Local Desktop Agent Reportedly Running on Kimi K2.6 With a 300 is a downloadable application for macOS and Windows that accepts plain-language goals and acts on your machine directly. You install it, describe a task, and it reads your local files, drives your real browser through a companion extension called WebBridge, and runs scheduled jobs on daily, hourly, or conditional triggers. The underlying model is Kimi K2.6, an open-weight Mixture-of-Experts architecture activating roughly 32 billion parameters per token across a 256K-token context window, with documented support for up to 4,000 coordinated steps inside a single run.

Operations teams feel this first

• Research analysts who manually pull data from gated portals every morning, now getting that output waiting for them on a schedule.
• Legal and compliance reviewers who spend hours cross-referencing local documents against live regulatory pages, without touching a single copy-paste.
• Ops managers who build no-code automations in tools like Zapier but hit walls the moment a task requires a real authenticated session or a local file path.

The people blocked by the cloud-sandbox ceiling are exactly who this is built for.

Local-first agents are pulling away from the hosted pack

Competitors including browser-based agent platforms have spent 2025 iterating on sandboxed remote sessions, a model that requires giving a third-party server access to your credentials. Kimi Work sidesteps that entirely by running on-device, which means the authentication problem and the file-access problem disappear together, and enterprise security teams have one fewer objection to field.

What it actually does on a Tuesday morning

• Monitor a live web portal and write findings to a local report on a timer.
• Search across hundreds of local documents in parallel using sub-agents.
• Fill forms in your real browser using your existing logged-in session.
• Split a large research task into parallel workstreams and merge the results.

Pricing is not listed publicly for Kimi Work at launch — check our directory for updates as they surface.

The honest tradeoff: running 300 sub-agents locally will stress consumer hardware, and performance on machines below a high-end threshold is not yet documented.

If local file access is not your bottleneck, browser-based agent tools may be faster to set up. If you need on-device scheduling without the swarm overhead, lighter desktop automation tools remain competitive for single-thread tasks.

The cloud-agent default is being challenged from both ends of the market

On-device AI agents are moving from proof-of-concept to production-grade faster than most enterprise procurement cycles can track. We cover tools like this every Friday — subscribe here and we’ll send the best ones straight to you.