I'm the maintainer of Google Workspace MCP (https://github.com/taylorwilsdon/google_workspace_mcp), a project that started more as a learning exercise than anything else but has grown into one of the most popular open source MCP servers out there today. It's been time consuming and at times overwhelming, but I can honestly say that it's also been one of the most fascinating periods of my 20+ year history of writing code for a living.
With the release of Workspace MCP v2, I put together the blog post linked here just looking back on just how much things have changed in what feels like a lifetime (but was really only the past year and a half). Coding harnesses have gone from a glimpse into the future that insisted on spewing out comments in code like they were going out of style, to daily driver tools with incredible capability in the right hands.
One of the single biggest challenges of maintaining an open source project in the age of AI, particularly one that's directly aimed at LLM usage and inherently attracts those most interested in AI is maintaining a quality bar and a code style while finding a balance between getting as many feature requests and PRs in as possible without ending up with a slop codebase that gets worse with every change. I still read every PR with my own eyes before merge, but the lions share of my workflow has turned into a two part adversarial code review with claude code & codex plus automated reviews from coderabbit in the repo.
In general, I don't use AI to write, whether it be my preferred medium of long winded prose, slack messages or hackernews posts. I absolutely do use AI to assemble the statistics in the article though! I also very much use AI to build Workspace MCP, and there is (what I think is) a pretty interesting chart in the article that shows the timeline of models and harnesses along the way.
Hopefully folks find this interesting, and as always, PRs welcome!
I'm the maintainer of Google Workspace MCP (https://github.com/taylorwilsdon/google_workspace_mcp), a project that started more as a learning exercise than anything else but has grown into one of the most popular open source MCP servers out there today. It's been time consuming and at times overwhelming, but I can honestly say that it's also been one of the most fascinating periods of my 20+ year history of writing code for a living.
With the release of Workspace MCP v2, I put together the blog post linked here just looking back on just how much things have changed in what feels like a lifetime (but was really only the past year and a half). Coding harnesses have gone from a glimpse into the future that insisted on spewing out comments in code like they were going out of style, to daily driver tools with incredible capability in the right hands.
One of the single biggest challenges of maintaining an open source project in the age of AI, particularly one that's directly aimed at LLM usage and inherently attracts those most interested in AI is maintaining a quality bar and a code style while finding a balance between getting as many feature requests and PRs in as possible without ending up with a slop codebase that gets worse with every change. I still read every PR with my own eyes before merge, but the lions share of my workflow has turned into a two part adversarial code review with claude code & codex plus automated reviews from coderabbit in the repo.
In general, I don't use AI to write, whether it be my preferred medium of long winded prose, slack messages or hackernews posts. I absolutely do use AI to assemble the statistics in the article though! I also very much use AI to build Workspace MCP, and there is (what I think is) a pretty interesting chart in the article that shows the timeline of models and harnesses along the way.
Hopefully folks find this interesting, and as always, PRs welcome!