This opens a shell where the docs are mounted as files. You can:
- tree to explore
- grep across pages
- cat specific files
The idea is simple: let agents read docs the same way developers read codebases.
Most coding errors from agents come from stale or incomplete context. Docs change constantly, but models are trained on old data. RAG helps, but it returns fragments. In practice, a lot of answers live across multiple pages or require exact structure that chunking loses.
Instead of retrieving snippets, this lets the agent just browse the docs directly.
Under the hood:
- we crawl the docs site and map each page to a file
- expose basic filesystem ops (read, grep, tree, find)
- run a lightweight bash-like shell client-side
- Everything is read-only and cached.
and write code against the latest docs instead of training data.
The reason for the filesystem approach is that agents already know it. Models have seen a huge amount of bash usage during training, so tools like grep and cat work out of the box without teaching new abstractions.
Right now this works on documentation sites. The longer term idea is making more of the web navigable like a codebase.
Would love feedback, especially from people building agent workflows.
I built agentsearch, a free tool that turns any documentation website into a browsable filesystem that you can access with one command.
npx nia-docs https://docs.anthropic.com
This opens a shell where the docs are mounted as files. You can:
- tree to explore - grep across pages - cat specific files
The idea is simple: let agents read docs the same way developers read codebases.
Most coding errors from agents come from stale or incomplete context. Docs change constantly, but models are trained on old data. RAG helps, but it returns fragments. In practice, a lot of answers live across multiple pages or require exact structure that chunking loses.
Instead of retrieving snippets, this lets the agent just browse the docs directly.
Under the hood:
- we crawl the docs site and map each page to a file - expose basic filesystem ops (read, grep, tree, find) - run a lightweight bash-like shell client-side - Everything is read-only and cached.
You can also plug it into agents in one line:
npx nia-docs setup https://docs.site.com | claude
Then the agent can do things like:
- grep -rl "webhook" . - cat getting-started.md
and write code against the latest docs instead of training data.
The reason for the filesystem approach is that agents already know it. Models have seen a huge amount of bash usage during training, so tools like grep and cat work out of the box without teaching new abstractions.
Right now this works on documentation sites. The longer term idea is making more of the web navigable like a codebase.
Would love feedback, especially from people building agent workflows.
Website: https://www.agentsearch.sh/