Small tools
that hold fast
A holdfast is the root-like grip that anchors kelp to rock. It is small, it is unglamorous, and the whole forest above it depends on it not letting go.
Computers got fast, and then we taught them to make us wait. Searching for a file that is right there. Watching a test suite tick. Agents spending their turns asking whether the job is done yet. We build the small pieces that take that waiting away.
How we build
Local first
Indexes and job records are stored locally. No accounts, no sign-in, and a daemon that only listens on localhost.
Measured, then claimed
Every speed number on this site comes from a benchmark you can rerun, with the method and raw data published next to it.
Agents are users too
JSON output, MCP tools and documented guidance for agents ship in the first release of a feature, not as an afterthought.
Durable by default
Crashes and restarts are normal. Indexes recover from their write-ahead log; job records, captured output and pending wakes persist in SQLite.
Small surface, deep roots
Each tool does one job. seekfs does not search contents; vanth is not a cluster scheduler. Saying no keeps them fast and dependable.
In the open
MIT and Apache-2.0 source, release notes for every version, and bugs written up honestly, with the numbers that found them.