I build an open-source local workflow creator so that you will always have a repeatable result
I kept noticing something while using AI agents: I was asking them to figure out the same tasks over and over.
The first time, that makes sense. The agent needs to understand what I want and work out how to do it. But by the tenth run, I already know the process. I just want it to follow those steps and finish.
Instead, it would plan again, make more tool calls, and sometimes take a different approach—even with the same prompt. The results were often useful, but the extra time and token usage started adding up.
So I thought, “Why don’t I let AI build the process once, then save that process as a workflow?”
That’s how I started building it.
I wanted to describe a task using my existing ChatGPT or Claude subscription, have AI turn it into connected steps, then review and reuse it. If something needs judgment, it can still go through AI. But if the next action is simply “check whether another page exists” or “save these rows to a CSV,” ordinary code can handle that.
It grew from there. I added conditions, loops, counters, delays, uploads, forms and API calls. You can watch the active step glow as it runs and see the output appear along the way.
Then I wanted to check whether it actually made a difference.
I ran a scraper benchmark with 10 paired runs. The prompt-driven MCP agent averaged 40.05 seconds, while the saved native workflow averaged 6.12 seconds—about 6.54× faster. Both finished every run successfully and produced the same expected results.
The agent averaged 153,304 model tokens per run, including cached input. The native workflow used zero model tokens during execution, because those particular steps ran entirely as code.
There’s a catch worth being clear about: this was a small test using fictional local pages and a prebuilt scraper. It didn’t include the cost of creating the workflow, and workflows that contain AI steps still use tokens. But it showed me that, for a predictable task, I didn’t need an agent to rethink the whole process every time.
I’ve made the source available so other people can try it too. It’s free for personal and noncommercial use.
So the link is hacksider/MCP-Workflow: MCP Workflow Creator
2 points · 6 comments
Mentions
I build an open-source local workflow creator so that you will always have a repeatable result I kept noticing something while using AI agents: I was asking them to figure out the same tasks over and over. The first time, that makes sense. The agent needs to understand what I want and work out how to do it. But by the tenth run, I already know the process. I just want it to follow those steps and finish. Instead, it would plan again, make more tool calls, and sometimes take a different approach—even with the same prompt. The results were often useful, but the extra time and token usage started adding up. So I thought, “Why don’t I let AI build the process once, then save that process as a workflow?” That’s how I started building it. I wanted to describe a task using my existing ChatGPT or Claude subscription, have AI turn it into connected steps, then review and reuse it. If something needs judgment, it can still go through AI. But if the next action is simply “check whether another page exists” or “save these rows to a CSV,” ordinary code can handle that. It grew from there. I added conditions, loops, counters, delays, uploads, forms and API calls. You can watch the active step glow as it runs and see the output appear along the way. Then I wanted to check whether it actually made a difference. I ran a scraper benchmark with 10 paired runs. The prompt-driven MCP agent averaged 40.05 seconds, while the saved native workflow averaged 6.12 seconds—about 6.54× faster. Both finished every run successfully and produced the same expected results. The agent averaged 153,304 model tokens per run, including cached input. The native workflow used zero model tokens during execution, because those particular steps ran entirely as code. There’s a catch worth being clear about: this was a small test using fictional local pages and a prebuilt scraper. It didn’t include the cost of creating the workflow, and workflows that contain AI steps still use tokens. But it showed me that, for a predictable task, I didn’t need an agent to rethink the whole process every time. I’ve made the source available so other people can try it too. It’s free for personal and noncommercial use. So the link is hacksider/MCP-Workflow: MCP Workflow Creator