DeepPCB

DeepPCB is a cloud routing engine built on reinforcement learning research. You upload a placed board file, or route inside KiCad with its plugin, and the engine routes the remaining nets to your rules, returning the board into the same design flow it came from. Engineers keep the nets that need their judgment and hand the routine routing volume to the engine.

It accepts Altium (.pcbdoc), KiCad with a plugin, EasyEDA PRO and STD, Zuken, and OrCAD, Allegro, and Eagle via Specctra DSN. A setup step called Cooper reads the board and reports what it understood, including stackup, net classes, differential pairs, planes, and clearances, and can be told in plain language to protect sections or mark diff pairs before routing.

The homepage states the engine can also be deployed in public cloud, private cloud, or on-prem, and integrated through a custom API and MCP or native plugins inside CAD tools. It is explicit about what it is not for: showcase routing, production layout where yield and manufacturability set the bar, and the densest high-speed boards, where DDR, length matching, and RF constraints stay with the engineer.

A Hacker News user in the beta era reported exhausting the 3-project free limit because all three test boards failed with errors, including an unexplained error and a non-45-degree rotation rejection, and that errored projects counted toward the limit. The DeepPCB team replied that it had refreshed the user's credits and would stop counting errored projects.

Source, 2019-12