Very interesting project! I make an MCP for woodworking [0] that has a similar set of constraints: citing everything to real world materials, sourcing them from real suppliers. I also support electrical components, for designs like this raspberry pi arcade cabinet [1].
For context, I've currently got 579 registered users. By far, most of them are using it to design brand new projects or to customize our published plans, rather than converting their existing designs into orderable BOMs.
I wonder if you have considered building configurable libraries of designs and/or PCB design features into your MCP? That could allow for building with in-stock, least expensive components. And it opens your product up to tinkerers that wouldn't otherwise design a PCB in KiCad.
I am one such potential customer! I am restoring a classic Skee Ball machine from the early 90s, I want to make a little esp32 board to replace the vintage system board. If I could go from prompt to PO Box for a reasonable price I would certainly be interested.
Sawdust also tracks shop inventory and capabilities, so if you have lumber and fasteners and your table saw configured you can filter down to things you can build with what you have on hand. Could be interesting to do something similar here with e.g. minimum order quantity leftovers.
Also an opportunity to become a marketplace -- if I can design my replacement board then publish it and get a cut of its sales, you can smartly replace OOS or now-more-expensive components, and keep the board purchasable and configurable, solving lots of problems for the open source hardware space.
This is something we are working on right now. The MCPs main focus is going to be integrating with design tools, especially the ones that agents are using. Would be very interested in collaborating on a marketplace. Your project seems cool and has similar intentions!
Hola HN, Will and Johnny from ProvenMetal (https://provenmetal.com) – we’re automating the circuit board supply chain.
Ordering an assembled circuit board is a seriously messed up process.
Roughly…
upload design
chase down manufacturer for a quote
manufacturer emails a bunch of technical questions
you answer the technical questions
they ask more technical questions
you answer more technical questions
then you realize that components that were available at the start of the process are no longer there, so you have to find and source substitute parts
you consign them to the manufacturer and try to co-ordinate with them
deal with shipping delays of the consigned parts, they say they won’t start until all components are there
they start the build, come across some design issue, stop working until it’s resolved
you resolve the issue
they finish and send it to you
and then it gets stuck in customs / a natural disaster happens to the delivery truck
We automated these steps for our own orders, collected data on every error, and built that into our own manufacturing co-ordination system.
I’m a big believer that this time next year, most new circuit board designs will be AI generated. Designs are useless if they can’t be manufactured well.
So we’re connecting agents to manufactured electronics. Design agents should know what can actually be built, with what parts, by when.
A bit on how the sausage is made:
Integrate with both domestic and overseas manufacturers, integrate with their quoting models, use APIs where available, and auto-select based on lead time, price, and capability.
Use manufacturing capability data and historical design data, with heuristic based harnesses, to automate design review. The design agent takes this data to improve designs.
Integrate with component vendors (big and small) to have complete supply chain coverage of every component based on availability, lead-time, and price. With big vendors, we use APIs. With small vendors, we use browser agents.
Manufacturers have technical questions that add days to lead time. We use agents to automate email responses. With enough data, you can predict their questions, and give them answers to questions before they have them.
Every order has a digital thread of every single action related to an order. Each order is constantly monitored to pick up on issues before they blow up lead times.
We’ve done $280k worth of orders in our first 3 months to develop this model.
Limitations:
Rigid-flex & flex PCBs - quotes aren’t instant and need more engineering review.
Complex fabrication processes – HDI and RF boards need more engineering review.
Stuff will break. We launch fast, find issues, and fix them fast.
Congratulations, Will and Johnny! This is the future of ordering PCBs.
What do you expect to be most common mistakes agents would make when ordering through MCP? And how to avoid those?
Very interesting project! I make an MCP for woodworking [0] that has a similar set of constraints: citing everything to real world materials, sourcing them from real suppliers. I also support electrical components, for designs like this raspberry pi arcade cabinet [1].
For context, I've currently got 579 registered users. By far, most of them are using it to design brand new projects or to customize our published plans, rather than converting their existing designs into orderable BOMs.
I wonder if you have considered building configurable libraries of designs and/or PCB design features into your MCP? That could allow for building with in-stock, least expensive components. And it opens your product up to tinkerers that wouldn't otherwise design a PCB in KiCad.
I am one such potential customer! I am restoring a classic Skee Ball machine from the early 90s, I want to make a little esp32 board to replace the vintage system board. If I could go from prompt to PO Box for a reasonable price I would certainly be interested.
Sawdust also tracks shop inventory and capabilities, so if you have lumber and fasteners and your table saw configured you can filter down to things you can build with what you have on hand. Could be interesting to do something similar here with e.g. minimum order quantity leftovers.
Also an opportunity to become a marketplace -- if I can design my replacement board then publish it and get a cut of its sales, you can smartly replace OOS or now-more-expensive components, and keep the board purchasable and configurable, solving lots of problems for the open source hardware space.
[0] https://sawdust.diy
[1] https://app.sawdust.diy/share/3469b5e6-88f8-43f5-ade0-c373d3...
This is something we are working on right now. The MCPs main focus is going to be integrating with design tools, especially the ones that agents are using. Would be very interested in collaborating on a marketplace. Your project seems cool and has similar intentions!
pretty cool, does it check whether components are in stock
Yes, has access to all the vendor APIs and makes subs if not in stock
Hola HN, Will and Johnny from ProvenMetal (https://provenmetal.com) – we’re automating the circuit board supply chain.
Ordering an assembled circuit board is a seriously messed up process.
Roughly…
upload design chase down manufacturer for a quote manufacturer emails a bunch of technical questions you answer the technical questions they ask more technical questions you answer more technical questions then you realize that components that were available at the start of the process are no longer there, so you have to find and source substitute parts you consign them to the manufacturer and try to co-ordinate with them deal with shipping delays of the consigned parts, they say they won’t start until all components are there they start the build, come across some design issue, stop working until it’s resolved you resolve the issue they finish and send it to you and then it gets stuck in customs / a natural disaster happens to the delivery truck
We automated these steps for our own orders, collected data on every error, and built that into our own manufacturing co-ordination system.
I’m a big believer that this time next year, most new circuit board designs will be AI generated. Designs are useless if they can’t be manufactured well. So we’re connecting agents to manufactured electronics. Design agents should know what can actually be built, with what parts, by when.
A bit on how the sausage is made: Integrate with both domestic and overseas manufacturers, integrate with their quoting models, use APIs where available, and auto-select based on lead time, price, and capability. Use manufacturing capability data and historical design data, with heuristic based harnesses, to automate design review. The design agent takes this data to improve designs. Integrate with component vendors (big and small) to have complete supply chain coverage of every component based on availability, lead-time, and price. With big vendors, we use APIs. With small vendors, we use browser agents. Manufacturers have technical questions that add days to lead time. We use agents to automate email responses. With enough data, you can predict their questions, and give them answers to questions before they have them. Every order has a digital thread of every single action related to an order. Each order is constantly monitored to pick up on issues before they blow up lead times.
We’ve done $280k worth of orders in our first 3 months to develop this model.
Limitations: Rigid-flex & flex PCBs - quotes aren’t instant and need more engineering review. Complex fabrication processes – HDI and RF boards need more engineering review. Stuff will break. We launch fast, find issues, and fix them fast.
Install MCP: https://provenmetal.com/mcp
Normal ordering flow: https://central.provenmetal.com/order
We are very interested in your opinion. This is a very messy problem to solve. Try out the MCP for free. Let us know what breaks.