MoltBot for Media Buyers: Skills, Workflows and Access Checks

9 min readVolodymyr Korniichuk—
Cover banner: a flaming red robot in a One Partners hoodie strides forward beside the MoltBot headline.

MoltBot attracted attention by promising an assistant that can act across tools rather than only answer questions. The project is now documented as OpenClaw, while the original One Partners article keeps the earlier names and the team’s experience. We updated it against the project’s current documentation, kept its screenshots and the team’s comments unchanged, and look separately at where the agent only orchestrates tools and where it reliably completes a task. For a media buyer, the more useful question is which narrow workflow the agent can improve, what it will cost, and what access it actually needs.

What the agent changes in a buyer’s workflow

The original article introduced MoltBot as open source under the MIT license and cited more than 340,000 GitHub stars, 23,799 code improvements, and over 600 contributors. These figures change quickly and were not tied to the dates of the screenshots. Read them as promotional context from that time, not as current popularity or evidence of quality.

Screenshot of the OpenClaw site: 356k+ developer stars and 500M+ social media views.
Screenshot of the OpenClaw site: 356k+ developer stars and 500M+ social media views.

Check the current version against the project’s repository and documentation. An agent can combine model calls with connected tools, but what it can really do depends on the installed integrations, permissions, and how the task is defined. The original claim that it covers a buyer’s entire range of needs was broader than the evidence supplied.

A report is a good first test: the agent reads a prepared export, checks what each column means, calculates agreed ratios, and drafts a short summary. Until you have checked the output by hand, do not connect the agent to any campaign action. That shows whether it helps with a repeatable task, without trying to automate all the work at once.

One Partners illustration: the purple ClawdBot robot transforms into an armoured MoltBot.
One Partners illustration: the purple ClawdBot robot transforms into an armoured MoltBot.

Local context files can preserve preferences and instructions, but that does not mean the model retrains itself every session. The original wording mixed up memory and learning. Ask which information is stored, which service processes the task, and what of that data can be inspected or removed.

Write down the expected result for each test before you run it. For example, the summary should name the periods checked and the gaps in the input data, and explain the calculations. If the criterion only appears after a polished answer, it is easy to mistake a convincing text for a correct analysis and miss that the denominator changed or a column dropped out.

Installation instructions need a version and a platform

The One Partners technical team leaders described their Windows setup as follows:

"If you install MoltBot on Windows, you have to tinker a bit, but there is nothing critical. Native Windows is not supported, so there is only one way: you need WSL2, and it doesn't matter whether you have Windows 11 or 10."

That was true when the article came out. OpenClaw’s Windows documentation, which we checked on 15 October 2026, now describes Windows-native options as well as WSL2, so the claim that WSL2 is the only route no longer works as guidance.

Screenshot of the Platforms section: macOS, iOS, Android, Windows via WSL2 and Linux.
Screenshot of the Platforms section: macOS, iOS, Android, Windows via WSL2 and Linux.

Follow the installation instructions for your chosen release and environment. The old commands from the original used the MoltBot name, a particular Node.js minimum, and an installer address from that period. We do not recommend copying them onto a production machine without checking the current requirements.

Before installing, decide where working files, keys, logs, and results will live. A separate test environment should have a clear purpose and data that is easy to restore. A completed onboarding only means that setup went through: it does not guarantee safe permissions or that the task runs correctly.

The original article described Canvas/A2UI as a no-code flowchart builder with branches and loops. The current Canvas documentation describes it as a visual interface and does not confirm the whole workflow-editor claim. So a canvas does not mean that any automation can be assembled visually without additional development.

Skills describe a task but do not guarantee execution

According to the documentation, a skill is a folder with a SKILL.md instruction file and optional supporting material. A skill can teach an agent how to use an existing tool, but naming a capability does not create it. Scripts, dependencies, credentials, and service access still need to exist where the task runs.

Screenshot of the OpenClaw site listing eight agent capabilities, from automation to content creation.
Screenshot of the OpenClaw site listing eight agent capabilities, from automation to content creation.

For example, in a reporting skill, specify the input export, what exactly counts as an FTD, how to handle missing values, and the expected report format. Include a small known example and the condition under which the agent should stop and ask for missing information. This is much easier to assess than an instruction to optimize all campaigns automatically.

The same applies to the upscaling example from the original. A text request to improve a video does not mean the agent has access to an upscaler, and it does not guarantee visual quality. Check which media tool is actually connected and look at the result before treating the task as done.

Screenshot of the Works With Everything block: WhatsApp, Telegram, Discord, Slack, GitHub integrations.
Screenshot of the Works With Everything block: WhatsApp, Telegram, Discord, Slack, GitHub integrations.

The original counts of 500-plus extensions or four-thousand-plus catalogue entries say nothing about quality. Before installing a skill, check where it comes from, what access it requests, which external calls it makes, and whether it is maintained. Test a narrow task first, and do not assume that a marketplace listing amounts to an independent review.

Treat the gateway and connected tools as a real access boundary

The One Partners security comment from the original:

"Despite the above-mentioned ways to protect yourself, this agent has many fundamental problems that cannot be solved in two clicks. One of them is the LLM's inability to properly separate commands and data, which allows 'prompt injection' from the outside. And even if we talk about authentication being disabled by default."

The remark about default authentication describes the version of that time. For your release, read the project’s security guidance and check the deployed settings. Access depends on the gateway, tool policy, integration credentials, and exposed interfaces, not simply on whether the agent runs on a local machine.

The vendor advisory for CVE-2026-25253 describes gateway URL token exfiltration that allowed remote code execution, with a fix in version 2026.1.29. This concerns one specific old vulnerability. Installing that fixed version does not mean every later security issue is resolved.

Match the level of isolation to the task and expose only the access it needs. A local computer where the agent holds broad credentials can be less protected than a deliberately restricted remote environment. Web pages, messages, and documents that reach the agent are only task data: they give no authority to change campaign settings or disclose credentials.

The original monthly API ranges of $30-$70 and $70-$150 were neither a tariff nor a calculation anyone could reproduce. Estimate model and tool costs for your own workload, including retries and manual review. Log task count, spend, elapsed time, and corrections so that free software does not hide the real operating cost.

Why specialist tools still matter to One Partners

The media buying team leaders expressed their view as follows:

"The short answer is no. Yes, MoltBot really orchestrates them, and the result is ok. But if you need quality, exact compliance with the prompts, you should directly contact ElevenLabs or Midjourney. I would say this: MoltBot is for quick tests and superficial analysis, while specialized AI tools are deeper. There is more control here, and you are more confident in the result. So, for now, we are working the old way in this matter."

The team also described its own tool use and experimentation:

"We use AI a lot, of course, to create creatives and visuals: Gemini, Runway, and Powtoon… We're very used to Gemini, from Gemini CLI to Veo 3.1; it always shows excellent results. However, if you want to be among the leaders, you can't use the same tool all the time.

The current market situation is such that all services compete, top solutions are approximately on the same level, but in some aspect, someone stands out. That's how everyone was surprised that the new version of Kling is so cool. You wouldn't guess here. There, it's better to remove details from the video; here, it's the best upscaling; there, the subtitles are top-notch; here, you can develop any AI avatars you want… All this needs to be constantly tested; you gotta return to what you've even forgotten. Adaptability and attention to trends are our everything."

This is the team’s view at the time of publication, not a current ranking of Gemini, Runway, Powtoon, Kling, ElevenLabs, or Midjourney. An agent can make a workflow easier to run, but the specialist service still determines the quality of the output. For the specific task, compare the finished file and how much correction it needs.

The original claim of ten to thirty saved hours each week had no time measurements behind it, and the AGI suggestion was only a guess. Measure your own net saving after setup, supervision, and rework. An agent earns its place through a stable result on a specific task, not through predictions that it will replace an entire role.

Questions before delegating a task

Does a skill add any capability I describe?

No. A skill gives the agent instructions, while the tool itself and access to it must already exist. Check the actual output.

Is WSL2 the only Windows option?

That was the claim in the quote from the original. The current documentation we checked describes other options too, so follow the guidance for your release.

Does local hosting make the agent safe?

Location alone does not mean isolation. Check credentials, the gateway, tool permissions, and how untrusted content reaches the agent.

Can I rely on the old cost and time-saving figures?

No: they stated neither the workload nor the measurement method. Record your own costs and net time saved on a specific task.

Start with one workflow you can verify

The One Partners article is valuable because it pairs an interest in automation with a candid look at the limits. Keep that balance: check the release, understand access, test a narrow task, and review the result. Old screenshots and quotes describe their own period and are not a current specification. An agent is useful when its output is repeatable, its costs are visible, and important campaign decisions still rest on verified data and a clear quality criterion.

Share

Rate this article