You’re Still Chatting With AI. Other People Are Putting It to Work.

For two years, everyone’s been teaching you how to write prompts. And what did you get? ChatGPT writes your copy, Claude fixes your code, Midjourney makes your images — but you’re still the one clicking every button.

Doing 3D work in Blender? You’re dragging the nodes. Editing photos in GIMP? You’re tweaking the layers. Cutting video in Kdenlive? You’re scrubbing the timeline yourself. AI is just a mouthpiece. The actual labor sits squarely on your shoulders.

That’s exactly why 99% of people fail at AI side hustles: the “thinking” AI does for you is worthless. The money is in the “doing” — and the doing was never automated.

That just changed.

HKU Open-Sourced CLI-Anything: AI Finally Has Hands

The HKUDS team at the University of Hong Kong released a project called CLI-Anything (repo: https://github.com/HKUDS/CLI-Anything, Apache-2.0 license, 48k+ stars). The idea is brutally simple: turn professional software into command-line tools that AI can call directly.

What does that mean? Think of it this way: every piece of software now has a phone number AI can dial. Before, if AI wanted to operate Blender, it had to screenshot the UI and simulate mouse clicks — slow, dumb, error-prone. Now AI just “dials in” a command like blender render --scene kitchen --output png, and Blender renders a kitchen scene on its own. One command = one action. No screenshots, no clicking, pure CLI execution. Stable and reproducible.

Confirmed supported software so far:

• Blender (3D modeling, rendering)
• GIMP (image editing)
• FreeCAD (engineering CAD)
• Kdenlive (video editing)

Three more things make it lethal:

1. Auto-generates the CLI, tests, and docs — you write zero adapter code
2. Structured JSON output — agents can pass data to each other and chain into pipelines
3. Confirmed working with Claude Code, Codex, OpenClaw, and Cursor

Bottom line: AI agents just grew hands.

What This Means If You’re Here to Make Money

What’s the freelance market at its core? Trading skill for money, trading hours for deliverables.

Search “3D modeling” on Fiverr or Upwork — gigs run $30-300. Video editing? A talking-head clip goes for $20-80. Logo design, photo retouching — anything GIMP can handle pulls $15-60 per order. What do these gigs share? Highly repetitive workflows, templated client requests.

Before, you could handle maybe 3 orders a day. Your hands were the ceiling. Now picture this workflow:

Client sends a brief → Claude Code interprets it → CLI-Anything drives Blender/GIMP/Kdenlive → finished product comes out → you just do QC and polish.

Your role shifts from “worker” to “foreman.” One person, output of a small studio.

The 3-Step Rollout

Full disclosure: Step 1 requires basic command-line skills (opening a terminal, pasting commands, reading error messages). If you’ve never touched a terminal, start with ready-made AI tools (Canva AI, CapCut) to validate demand first, work through a CLI beginner tutorial, then come back to this stack.

Step 1: Set up the environment (1-2 days)

Repo: https://github.com/HKUDS/CLI-Anything. Per the official README, the community CLI installs like this:

1
2

pip install cli-anything-hub
cli-hub install <software>   # e.g. cli-hub install blender

Requires Python ≥ 3.10. Once the CLI for your software is installed, hook it into Claude Code or Cursor and run the full loop: “one natural-language sentence → software executes automatically.” Practice on 10 test briefs first and lock in prompt templates for common requests.

Step 2: Open shop (week 1)

Your main battlefield: Fiverr and Upwork — zero barrier, massive traffic, mature categories for digital services. Titles should hit the pain point directly: “Video editing, 24-hour delivery.” “3D modeling & rendering, same-day turnaround.” On pricing, sell speed, not cheapness — clients pay a premium for same-day delivery, while rock-bottom prices only attract the pickiest buyers (more on that in the risk section). In parallel, post screen recordings of “AI editing a video start to finish” on TikTok and Lemon8 — this kind of “watch AI do the work” content is a traffic magnet on its own, funneling viewers into higher-ticket private orders.

Step 3: Scale and clone (weeks 2-4)

Once orders flow in, wire the pipeline together with n8n — a no-code drag-and-drop automation tool that connects “order comes in → AI does the work → delivery notification” like Lego bricks. Don’t know n8n? Fine. The semi-manual mode — human takes orders, AI does the work — is more than enough early on. Learn it when volume forces you to. Once one vertical works, clone sideways: editing works → add modeling → add photo retouching.

Let’s Do the Math (With Assumptions, Not Promises)

Take video editing gigs. Everything below is hypothetical modeling, not an income guarantee:

• Price per clip: $45
• Manual era: 3 clips/day → $135/day
• Agent era (ideal scheduling): 15 clips/day, 10 min QC each → $675/day, minus ~$40-60 in API costs → $600+ net

But that’s the rosy case. Three realistic tiers:

• Conservative (first 1-2 months): New shop, no traffic. First two weeks might be zero orders — you’ll grind cheap gigs and reviews to build ranking. Assume 2-3 orders/day, 30% revision rate (client rejects the AI output), $80-120/day → roughly $2,400-3,600/month.
• Base case (established shop traffic): 6-8 orders/day, 20% revision rate, $250-350/day → around $8-10k/month.
• Optimistic (multi-platform + private high-ticket clients): Fiverr + TikTok funnel + agency subcontracting — that’s what it takes to touch $15k/month.

The real variables are client acquisition and revision rate, not AI speed. AI compresses production time. Finding clients, clarifying requirements, handling revisions, after-sales — none of that shrinks by a minute.

Risk Warnings (More Important Than the Money Part)

1. The “wait, AI made this?” problem: Clients who discover AI did the work may demand refunds or haggle. Sort out your positioning early — you sell “results + speed,” you own the QC standard, and every AI draft gets a human pass before delivery.
2. Platform policy risk: Marketplaces may tighten rules on AI-generated deliverables. For commercial orders, nail down copyright ownership with the client before taking the gig.
3. Don’t fight a pure price war: Once automation spreads, cheap is something everyone can do — and the gold rush dies in the race to the bottom. Differentiate on speed and reliable delivery.
4. Tool risk itself: This article is based on CLI-Anything’s official public docs (the README shows 2,461 passing tests), but I haven’t personally tested every supported app. Validate with 10 test briefs before spending a dime on shop setup — not the other way around.

The Window Won’t Stay Open Long

Every tool revolution pays the first people who turn the tool into an assembly line. 2023 was AI art commissions. 2024 was AI ghostwriting. 2026’s answer is sitting right here: AI directly operating professional software for automated delivery.

CLI-Anything is iterating fast (repo News updated through late May 2026, with new software integrations landing almost weekly), and few people know how to monetize it yet. But projects like this spread in weeks. By the time tutorials are everywhere, the competition won’t be about the tool — it’ll be about operations.

Do one thing today: open https://github.com/HKUDS/CLI-Anything, set up the environment, and get your first automated Blender render running tonight. If it works, then we talk about opening a shop. If it doesn’t, you’ve lost nothing but one evening.