# The Agent-Drives-Your-App Kit
OpenClaw is an open-source AI assistant (openclaw/openclaw on GitHub, 386,308 stars measured on
2026-08-14). A published demonstration shows it operating Blender directly: recolouring a 3D
character, moving its eyes further apart, enlarging its arms, animating it, then building a page for
it and deploying that page to a public link opened live on a phone. Nobody in that demo touches a
modelling tool. They type.
The Agent-Drives-Your-App Kit is a free three-file reference: plain markdown, no install, no
account. It covers the prompting method the demo actually uses, the install path with a sane opening
security posture, and an honest account of how an agent reaches an application at all.
## The prompt is a checklist, not an incantation
This is the part worth stealing, and it is deliberately boring.
The instinct with a capable agent is to write a paragraph: *make me a cool little robot character,
something friendly, kind of retro.* That gives the agent nothing to check itself against. You get one
interpretation and no way to steer except by writing another paragraph.
What the demonstration actually types is a shopping list. Dark oval eyes with bright cyan emissive
pupils. One curved organic antenna. Rounded arm nubs. Beveled leg stubs. Dark background, five-light
setup. One property per line, and every line is something you can tick or cross by looking at the
result.
That difference is the whole reason it works. A checklist gives each line a pass or a fail, so the
agent can verify its own work. When one line comes back wrong you correct *that line*, instead
of re-describing the entire object.
One change per request, and let them stack
After the first pass, the demo does not send a new checklist. It sends one line at a time.
*Make it more red.* Then: *move the eyes slightly further apart, and make the arms a little bigger.*
Then: *now animate it, I want it to wave.*
Each request lands on the state the previous one produced. Nothing is rebuilt from scratch, so the
work accumulates rather than resetting. That accumulation is the difference between an agent that
behaves like a slot machine and one that behaves like a collaborator, and it is the single habit most
worth copying out of the whole demonstration.
The failure modes are just as instructive. If the agent rebuilds everything instead of editing, the
request re-described the whole object rather than naming a delta. If you cannot tell whether it
worked, the properties were not checkable. *Make it look professional* has no pass/fail; *five-light
setup, dark background* does. If it keeps drifting back to an earlier state, the thing that must not
move was never written down as a constraint.
Two things the clips are leaving out
There is no Blender feature in OpenClaw. Search the repository and you get nothing:
gh api search/code -f q="blender repo:openclaw/openclaw" --jq '.total_count'
# -> 0
The agent reaches the application one of two ways, and both are real. It can go through a
tool/skill/plugin bridge (usually MCP) that exposes the application's own scripting API as callable
tools; the repository carries 274 files on MCP paths, and Blender has a Python API built for exactly
this. Or it can drive the graphical interface the way a person does: the project merged an embedded
macOS computer-control provider on 2026-08-14. The first is more reliable, because the agent
manipulates real objects instead of pixels. Neither is a button you press.
386,308 is a star count, not an install count. A GitHub star is a bookmark; the API exposes no
usage telemetry whatsoever. What the number does support is scale of attention: it is the most-starred
AI agent on GitHub by a wide margin, and the sixth most-starred repository of any kind. The nearest
agent-shaped repositories sit around 230,000 to 272,000, and every repository above it is a curated
list or a learning-resource collection rather than software.
Before you point it at anything
From the project's own documentation, and it deserves more attention than it usually gets: tools run
on the host for the main session unless you configure sandboxing. An agent that can operate
applications can operate yours, on your files. Inbound messages should be treated as untrusted input,
and channels that accept direct messages pair unknown senders by default.
A sane opening posture is local only, sandboxing on before any channel is connected, one channel at a
time with senders paired by hand, and a scratch directory rather than the only copy of something you
care about. A third-party bridge typically executes arbitrary code inside the target application.
That is the feature and the risk in the same sentence, so read what it runs, check it is still
maintained, and try it against a file you are willing to lose.
What I did not test
I did not run a Blender bridge end to end. I confirmed both mechanisms exist and confirmed there is no
Blender code in the repository itself, and the kit's bridge page tells you what to evaluate rather
than walking you through a setup I have not completed. A guide that sounds confident about an
integration nobody tested costs more trust than it buys.
The half with no dependencies needs no agent at all: the checklist prompt works with whatever
assistant you already use, on whatever application you already have, starting now.
The part that does not expire
Every specific above rots on its own clock. The star count is stale within days. The install line
changes with a release. A bridge that works today may be unmaintained by spring.
The checklist method does not rot. It is a way of writing a request so an agent can act on it and you
can verify the result, and it transfers unchanged to the next model, the next agent and the next
application. If you take one thing, take the shape rather than the tool.
Comment CLAW on the video and the kit comes to you directly.