Deepseek Harness 2.0 (Desktop App, Creator Mode): RIP Codex! The BEST OPEN Harness YET!?
In this video, I'll set up the new DeepSeek Harness desktop app (v0.2.0 release candidate 2) and actually use it on a small project instead of just reading the announcement. I give it a fictional merch shop with two broken sales calculations, have it repair the summarizer without touching the CSV or the tests, turn the results into a self-contained HTML report, and then add a product filter through a follow-up conversation. After that, I build a small Report checklist plugin with Creator mode, schedule a one-time automation task, and explain the Standard, PTC, Minimal and Creator modes. Finally, I go over the model settings and what DeepSeek V4.1 Flash costs through the API. -- Key Takeaways: 🚀 DeepSeek Harness now has an official desktop app for Apple Silicon Macs and 64-bit Windows, and the packaged app carries its own runtime, so you don't need Node just to use it. 🐧 On Linux, the documented route is the web interface: npx @deepseek-ai/dsh web, which needs Node 22.19 or later in the 22 series, or Node 24 and above. ⚙️ Turn on Show coding view in General settings to get the coding trajectory, code changes and agent presets, and remember that adding a workspace and selecting it are separate steps. 🔧 On the Northstar demo project, Harness skipped cancelled orders and used the real refunded units, all four original tests passed, and net revenue went from $977 to the correct $662. 📊 In the same conversation, it generated a self-contained report.html, a build_report.py generator with a --check option, and calculations.md, all reusing the repaired calculation. 🎯 The follow-up product filter showed Launch Tee at $200 and 8 net units, kept the whole-shop totals clearly labelled, and Reset worked inside the real Harness preview. 🧩 Creator mode produced a small Report checklist plugin that installed from a local folder, appeared in the sidebar without a restart, and reached 3 of 3 before Reset. ⏰ The Automation tasks plugin ran a one-time reminder that wrote the verified totals to a file, but a delivery record only means the instruction reached the conversation, so check the actual output. 💸 Harness is MIT-licensed, but model calls are billed separately: V4.1 Flash was 15 cents per million uncached input tokens and 60 cents per million output tokens off-peak, and one prompt can turn into many requests. ⚠️ It needed steering on both the report and the plugin, so keep tasks specific and start with a project where you already know what a correct result looks like.
