Command Code's desktop app, tested: a $1 coding agent with a plan-review loop, on DeepSeek V4 Flash

Install

Downloads exist for macOS (DMG), Windows (installer) and Linux (per-distribution packages). The agent runtime is bundled, so you don't need to install Command Code's CLI first. Add a project with Add project → Open folder; he used a small disposable repo, which is the right way to trial any new agent.

Layout: projects and chats on the left, the conversation in the middle, and a workbench on the right with Files, Changes, Browser and Plan views. The terminal opens as a drawer under the chat, so a dev server can keep running while you review.

Model and effort

The composer's model picker groups models by provider — DeepSeek, GLM, Kimi and Qwen among them — with search and favourites. The test used DeepSeek V4 Flash at max effort; default and high are also offered. Current rates for these models are in our calculator.

The plan-review loop — the part worth copying

In plan mode he asked for search and status filters on a task list, with a clear finish line: case-insensitive search, working together with the status filter, an empty state, and a stated test plan. The plan opens in its own panel with three choices: Accept and build, Keep refining, or Build without asking.

He used Keep refining, because the first plan had the substring comparison written backwards and was vague about clearing filters. The revision came back as version two with the changes highlighted — a proper place to catch a wrong approach before it becomes a diff.

What actually happened during the build

• The first test command couldn't find Node in the agent's shell. It located an installed version and retried with an absolute path. A desktop app doesn't remove local environment problems.
• It noticed it had forgotten the "all statuses" option and fixed that unprompted.
• All 11 tests passed, which he confirmed independently.
• The agent's browser-automation tool wasn't installed, so he did the interactive checks himself in the app's browser panel.

Command Code also documents a tool-call repair step: it validates the arguments a model sends to a tool and repairs certain malformed calls rather than failing. That is exactly the weakness of cheaper models in agent loops, so it's a sensible thing for a budget-model harness to invest in.

Design mode and review

With the dev page open in the browser panel, design mode lets you click an element and attach its context to the chat, so you can say "make this control's active state clearer" without naming the component. The docs also describe /design checkup (a broader review), /design smell (a report on generic design choices) and /design deslop (apply changes) — he did not run those, so treat them as documented rather than tested.

Before committing, open Changes: the diff also includes the "taste" files the app writes as it learns your preferences from what you accept, reject and edit. Review the whole list.

Pricing, as quoted in the video

• Go: $1 a month plus a processing fee, including $10 of monthly credits.
• $10 plan: $10 a month plus the fee, advertised with up to $70 of usage through model allowances.
• Allowances draw from one shared plan — not a separate full budget per model — so check the allowance for a model before assuming it costs the same.

Plan prices change; confirm on Command Code's pricing page before you rely on these.

Key Takeaways

• Use Keep refining on the plan — it caught a real logic error before any code was written.
• Give the agent a finish line: how features interact, the empty state, and how it will be tested.
• Trial any new agent in a disposable repo, and read the full diff including files the tool writes for itself.
• Cheap models plus a harness that repairs malformed tool calls can cover a lot of everyday work — his conclusion is "try it before paying for a $200 plan", not "cancel it".
• It's an alpha: expect environment rough edges like the missing-Node issue.