Jev tooling, a dependable AI harness, API integrations and a few tools of my own.
Jev tooling
Every request and answer, live, with a clear sign when Jev goes quiet.
Whole sample pages from a short brief, each one graded by Jev.
A fair head to head test, and the better tool kept for each job.
Tools that run on Jev or a local model, with nothing breaking.
Say a page or style in everyday words and Jev finds it.
Type a sentence and an AI works the browser or the Mac to get it done.
The harness
Every finished job comes with evidence that it was checked.
Work begins almost at once, with no long wait.
A broken login is reported at once, so nobody works from bad data.
Mistakes are caught early, so what you receive is already reviewed.
API integrations
Pasted code becomes real Webflow elements in the site’s own styles.
The login renews itself, even when Webflow asks it to slow down.
A loose request finds the right page, or becomes a question.
The chat keeps answering, whichever AI provider has room.
A small open model does real building work beside me and Claude.
People plan the day together and see each other’s changes live.
Own tools
Contact
A live view of every request sent to Jev and every answer returned.
I specified the behaviour and the commands, then tested and corrected it. Built with Claude Code, which wrote the code.
I can watch Jev work in real time. A call that gets no answer is recorded, not lost.
It runs on a single Mac and has no public demo.
A page builder where Jev judges the result and never writes. The pages shown are samples for fictional brands.
I set the rule that Jev never writes, and improved the output when early pages looked generic. Built with Claude Code, which wrote the builder.
A short brief becomes a finished page that Jev has graded.
No outside reviewer has rated the pages.
A test of Jev against the tool I already used to sort my own messages.
I designed the test before adopting Jev. Built with Claude Code, which wrote and ran it.
Jev is kept for judging a choice. My own tool is kept for sorting my messages.
It was a single test on my own messages, so it says little about other jobs.
A single switch that points my tools at Jev or at an open model that runs on my own Mac.
I directed the switch and the local setup, built with Claude Code. Credit to convaiinnovations for the open model and mizorewww for its Apple Core ML version.
The switch moves my tools between Jev and the local model. Nothing breaks when I switch.
I do not say it beats Jev, because the two were never tested on the same task.
Helpers that let Jev find the correct page or style on a large live Webflow site.
I set the rule that a choice between existing things goes to Jev. Built with Claude Code, which wrote the helpers.
On my test requests Jev outperformed keyword matching and made no wrong changes.
The test set is small and I wrote it myself.
Agents that work in a browser or on a Mac, with Jev confirming at the end that the goal was met.
I connected open source agents to Jev, with Claude Code. Credit to the browser-use and open-computer-use projects, which wrote the agents.
A task counts as done only when Jev confirms the result matches the goal.
These are demos, not a measured success rate.
An AI that has to show proof before it says it is finished.
I decide what counts as done and direct the work. Built with Claude Code.
Every finished job comes with evidence that it was checked. You are never asked to take the AI’s word for it.
A fast start for the whole system.
I direct the work and test the result. Built with Claude Code.
Work begins almost at once instead of after a long wait.
Safeguards on the connections to outside services.
I direct the work and test the connections. Built with Claude Code.
A broken login is reported immediately as a clear error. Nobody works from bad data hidden behind a number that looks normal.
Reviews built into the AI’s routine.
I direct the work and review the results. Built with Claude Code.
Many mistakes are caught early, before they reach you. What you receive has already been reviewed.
Turns pasted HTML and CSS into native Webflow elements that follow the site’s own styles.
I set the rules it works under, including never overwriting an existing style, and tested it in my own Webflow Designer. Built with Claude Code.
Pasted code becomes native Webflow elements that reuse the site’s styles. Pasting a newer version replaces the old block.
Proven on test pages. Live use has only just begun.
A standalone program that signs in to Webflow’s AI connection and stays signed in.
I own the workflow it serves and the Webflow account it signs in to. Built with Claude Code, which wrote and ran the code.
The login renews itself, and the program keeps working when Webflow asks it to slow down.
Tested offline so far. A completed live sign in is not yet on record.
A careful way of working through Webflow’s AI connection.
I set the working rules, review every change and publish by hand. Built with Claude Code.
A loose request resolves to the right page, or becomes a question when Jev is unsure. Publishing stays a manual step.
The checks here are small, and I wrote them.
A chat that keeps working across many AI providers.
I chose which providers to use and own the accounts. Built with Claude Code, together with the Chat Studio app that uses it.
A request still gets an answer when a provider is busy or refuses it. When none can answer, it reports why.
No automated tests yet. Confidence rests on live use.
A local app where I, a small open model and Claude share rooms and build together.
I own it and use it myself. Built with Claude Code.
The small model works safely inside its own workspace, and every job keeps its full goal.
A local tool, not a product.
A shared day planner. People join a room by code.
I own the tool and used AI coding tools for part of it.
Everyone sees each other’s tasks update live, with drag and drop, day notes and mentions across rooms.
A personal tool, not a product.
An editor that turns a site’s details into a clean llms.txt file.
I designed and built it.
A finished file, ready to publish, with a live preview as you type.
A schema editor that writes valid JSON-LD for an organization as you type.
I designed and built it.
The finished schema is copied in a single click.
A shared day planner with a column for each person, split into morning and afternoon.
I own the tool and used AI coding tools for part of it.
Everyone sees the same plan, updated live.