The pitch is not that it searches more but that it searches better material — newsletters written by developers about things they built, rather than SEO pages restating documentation. Returning URLs instead of full text is deliberate and works in an agent's favour: the model triages descriptions cheaply and only fetches what it will actually use. Take the remote endpoint unless you have a reason not to; the local route's known Claude and uvx issues are documented precisely because they bite.
A newsletter-backed research server. Rather than searching the whole web, it searches a curated body of developer-written newsletters, articles and blog posts, and you can plug in your own favourite newsletter to widen the base. Results come back as a short description plus a URL.
- Topic research across high-quality, developer-curated content rather than generic search results
- A results shape built for agents: short description plus URL, so the model can then fetch the full article for the ones worth reading
- Your own newsletters added to the knowledge base to widen what is searchable
- Three ways to run it — a remote HTTP endpoint, a local stdio process, or a Docker container
The recommended path needs nothing installed: point the client at `https://devbrain.svenai.com/mcp` as an http server. Locally it runs through uvx as `uvx --python ">=3.10" --from devbrain devbrain-stdio-server`, so Python 3.10+ is required — the README warns that VS Code ships 3.9, which is why the version constraint is on the command. Claude Desktop needs `--force-reinstall` in the args to pick up the current version, and the README documents the known uv and uvx path problem with symlinks into `/usr/local/bin` as the fix. Docker is `svenai/mcp-devbrain-stdio:latest`. The package is `devbrain` (1.1.9 in pyproject). MIT licensed.