feat: add native Model Context Protocol (MCP) server support (--mcp) - #910

I'm not sure why this is necessary. How does it help?

Coverage was decreased. Can you add tests?

Thank you for the review! I've addressed all three points in the follow-up commit 060fb51.

1. "I'm not sure why this is necessary. How does it help?"

The goal is to enable AI coding assistants to fully automate the PCB-to-G-code workflow without manual steps.

A concrete use case: tools like KiCad-MCP-Server allow an LLM (e.g. Claude) to design a PCB directly in KiCad. Once the design is done, the same AI session can call pcb2gcode --mcp to generate G-code from the Gerber files — all without the user ever opening a terminal. This closes the loop from schematic → PCB layout → G-code in a single AI-driven session.

I tested this personally on Windows 11 + WSL2 (Ubuntu 24.04) using Antigravity IDE, successfully generating G-code from KiCad-exported Gerber files via the MCP interface.

2. "Coverage was decreased. Can you add tests?"

Added the following in commit 060fb51:

• tests/mcp_server_tests.cpp — 3 Boost unit test cases:
  • mcp_server_initialize_test: verifies initialize returns correct serverInfo
  • mcp_server_ping_test: verifies ping returns a valid JSON-RPC 2.0 response
  • mcp_server_tools_list_test: verifies tools/list returns exactly 3 tools
• tests/options_tests.cpp — added mcp_option test case to verify the --mcp CLI flag is accepted
• tests/CMakeLists.txt — registered mcp_server_tests as a CMake test target
  All tests pass (*** No errors detected) on Ubuntu 24.04 with GEOS 3.12.1 and Boost 1.83.

3. "Rather than including this file, can you add it to the cmake?"

Removed the vendored src/nlohmann/json.hpp (24,765 lines). Now handled in CMakeLists.txt:

find_package(nlohmann_json 3.11 QUIET)
if(NOT nlohmann_json_FOUND)
  FetchContent_Declare(nlohmann_json
    URL https://github.com/nlohmann/json/releases/download/v3.11.3/json.tar.xz)
  FetchContent_MakeAvailable(nlohmann_json)
endif()

Uses the system-installed version if available, otherwise downloads automatically. No manual installation required.

Thanks for those improvements!

I am on holiday now so I can't fully read them. I'm also not an expert on MCPs so I don't really know what they do...

Why can't AI just run pcb2gcode with the correct command line and generate the command line? I don't understand why an mcp is needed. This is not to say that you did it incorrectly! This is just because I don't know what an MCP is!

Anyway, I'll be able to look at it in about 10 days.

I'm understanding it better. It seems like a nice improvement. It also provides a nice foothold for the mcp. In the future, maybe there will be more mcp commands.

I have added another test to improve test coverage.

coverage: 60.874% (-3.4%) from 64.253% — nando256:master into pcb2gcode:master

I'm still working on it. I'll try to keep the majority of the coverage done from cpp tests and only do what I have to from the python test.

@nando256 I'm happy with the coverage now. All my changes are here. If you're good with them, I'm ready to merge this.

Any thoughts?

Thank you for the thoughtful improvements! I've reviewed your changes and they look great to me:

• Extracting the stream-redirect boilerplate into run_mcp_with_input() is a clean improvement — much easier to read.
• The tools/call tests using the real binary path (PCB2GCODE_TEST_BINARY) are a great addition, covering the fork/exec path that the unit tests couldn't reach.
• The end-to-end stdio shutdown test via Python is a nice integration layer on top.

I'm happy with all the changes. Please go ahead and merge whenever you're ready!