Have you struggled with outdated recommendations from your AI assistant and/or burning through tokens like they’re going out of style because it’s searching the web for 30 seconds every time you want to understand which scene.getMeshBy(*) is the best for your use case? Add this mcp server to your ai assistant so that you can keep moving quickly instead of having your AI coding tool google everything…. based on latest commits from docs and code instead of random (or outdated/deprecated) internet content.
Originally I created for myself and ran locally but realized this could be more generally useful to the community so put this up. Let me know what you think and/or if you have problems setting it up (claude was a little wonky, not sure about copilot or others, haven’t tried yet).
@PirateJC@Deltakosh sorta trying to up the AI game and avoid using outdated APIs/patterns…I have an idea on how y’all could influence AI coding agents to “not” use old patterns and give upgrade instructions/paths, but we’d need to rap on how that would work.
Also curious…are the forums api accessible by date? I kinda think baking those discussion into a "faster” RAG/vector DB would really help accelerate newbies. 1/4 baked solution right now hosted on a machine in my basement but would love to pressure test the overall usefulness… originally I was going to bake havok physics in (because it matters more to me) but I’m thinking I may pivot to the editor first as it actually might be more widely useful (and it turns out I didn’t really know about it until I watched one of the podcasts)
and if you’re not a developer and trying to “vibe code” something…you probably need this 10x more otherwise you will find yourself in a rabbithole with no bottom.
Very neat! Will have to try this out. What tools does it expose to MCP clients? It sounds the server is hosting an index of the BabylonJS source code, is that right?
yeah I indexed the source code, api docs, and documentation for babylonjs core in a vectordb. “when it works” it’s multiple times faster if you’re looking for api docs or implementation details (like claude, even when looking in my local node_modules can take a minute or more for some things and this returns results in a second or two). What I’m noticing is that it gives more headspace for the coding assistant to “find more useful stuff” in fewer passes and equivalent token usage.
Also not, this is not battle hardened…frankly I wrote it for myself to speed my development up and stuck it on a server in my homelab. You CAN, however easily self host if you find it useful and my server dies/is too slow.
What I’d “really” like to do is be able to flag “canonical” or “current” approaches as “recommended” to avoid fetching outdated/old approaches (like old physics/audio engine) that show up in general search results. Not sure how I’d accomplish unless it was in the documentation or github somewhere.
