By humans · For AI agents · To help humans

Engineered, not generated.

The name is literal: everything plus the kitchen sink, built for AI. Instruments for the work agents actually do, in Word, PowerPoint, Excel, and on the open web. Four shipped products, one bar: everything included, all of it verified. Every number on this page is measured by a script and enforced by a test; we would rather show than sell.

The products → Downloads, docs, and the full catalog

4Servers

6,442Tests

0Corruptions

1Patent pending

The products

The name is the spec. Every server ships with everything: the capability an agent needs today, the fifty it will reach for next quarter, and the rails that make it safe with real work. An F1 car for the daily commute: overbuilt on purpose, so the spec sheet does the selling.

KitchenSink4Wordv2.1.2

The benchmark for what an agent can do in a document. Complete enough that the comparison table writes itself, fast enough that nobody watches the clock, and engineered so the work survives every single time.

221 ops · 110 tools · 1,846 tests · 0 corruptions

PyPI GitHub Docs

KitchenSink4PPTv1.2.2

Decks at the speed of thought. The only server anywhere that turns an agent's drawings into native, hand-editable slides: polished output people can keep working on, with zero babysitting in between.

142 tools · 1,251 tests · 1,500+ adversarial calls · 0 corruptions

PyPI GitHub Docs

KitchenSink4XLv1.2.2

Spreadsheets an agent cannot silently break. Every number knows whether it is fresh, every edit carries its formulas, and the answers arrive at interactive speed. The trust layer Excel always needed.

129 ops · 69 tools · 1,294 tests · 0 corruptions

PyPI GitHub Docs

KitchenSink4Webv1.0.1

The web, read whole and priced honestly. An agent sees the entire page for a fraction of the cost, accounts for what it skipped, and acts only after a human allows it. Fast, complete, and built to ask before it acts.

2,066 tests · 7× cheaper reads · 0 acting tools until allowed

PyPI GitHub Docs

Research

Research · Patent pending

Didymoi

Ask a model the same question twice and you get two different answers. Didymoi makes AI-assisted analysis reproducible: a deterministic layer builds the record, and the model works inside it without the power to alter it. Grounding handed over as a fence, not a springboard.

A way of making a probabilistic system answerable to a deterministic one.

Open tooling · MIT + NC

claude-code-statusline

The context gauge for Claude Code: real burn rate, turns remaining, rate limits, and alerts that escalate before the wall, not after. Built after a session nearly died with the stock gauge reading fine.

GitHub

How we work

It has never been easier to ship a tool that almost works. Almost is the whole problem. We build the ones that just work.

R-01

Measured, not asserted

Every number comes off a measurement script, and the build fails the moment the code stops backing it. Marketing does not get a vote.

R-02

We break it first

Nothing ships off one good demo. It ships after we have attacked it deliberately, repeatedly, and on the record.

R-03

Done means verified

Running is not done. Verified is done. Backups first, receipts after, every time.

R-04

Honest over impressive

A confident wrong answer is the one thing we refuse to ship. When something cannot be done right, it says so, and says why.

R-05

Limits in writing

What we cannot do is written down. Skips name their reasons. Changelogs name what actually changed.

R-06

Built by its first customer

All of it runs daily in the founder's own work before anyone else touches it. The rough edges die at our desk, not yours.

The business

Open where it builds trust, licensed where it captures value. The servers are AGPL-3.0 and free under its terms, forever; that is the funnel. A company building them into a product of its own buys a commercial license; that is the revenue. The model is simple, already in writing on every repository, and it scales with every integration.

Reach the humans →