AtlasCase 01 / 07Studio product / AI workbench

VeritasAxis studio

Telling you who is on the other end

Cloaky is a pseudonymous AI workbench: a full agent harness — threads, diffs, subagents, permissions, git review — pointed at whatever intelligence you choose, from a model on your own machine to Venice AI to your own Claude subscription. The design problem was honesty: every session has to say who can see it, and what it cost.

Role
Founder — product, design, build
Period
2026
Scope
Product strategy / Positioning / Desktop app UX / Design system / Marketing site / Brand
Opt-inEvery backend, Claude included
Per sessionPrivacy badge + cost ledger
Base + ProFree build and full workbench

The problem

AI coding tools tend to describe one provider’s contract as a property of the whole product. Cloaky routes to local models, Venice’s private and anonymized tiers, coding plans, and identified paths like a Claude subscription — so a single “private” claim would be false for most of them.

I designed and built the product end to end: the positioning, the desktop workbench, the model picker, the per-session privacy badge and cost ledger, and the cloaky.dev site with its interactive product tour.

Working idea
Name who is on the other end — every session.

How it fits together

Four routes. Four different promises.

Explore the system

  1. Model turns stay on this machine; connected tools may still use the network.

    See it in the workLocal art direction

  2. Contract-enforced zero retention on Venice’s private tier.

    See it in the workExplicit engine choices

  3. Identity hidden upstream; prompt content visible — and labelled.

    See it in the workSession controls

  4. Your own subscription: a direct path, stated plainly.

    See it in the workPrivacy in the review view
The whole system

Key decisions

The choices that shaped the work.

01

Sort by privacy first, price second

The model picker orders options by what the provider can see, then by cost, and spends its warning colour only where a choice is weaker than the one asked for.

02

Badge the session, not the product

Each thread names what it actually got, and the ledger records what it cost per response rather than estimating. Privacy becomes visible state, not a settings page.

03

Fail closed instead of pretending

Encrypted turns cannot carry structured tool calls, so the harness refuses tool turns under encryption rather than let an agent narrate edits it never made.

04

Show the product working

The site leads with an interactive tour of the real workbench — agent, review, rewind, subagents, engines, bots — instead of abstract claims about it.

The work

01 / 08

  1. Overview. “Keep the workbench, choose the intelligence”: Solo and Squad modes as two faces of one tool, with every provider mark on the same bar.

Overview. “Keep the workbench, choose the intelligence”: Solo and Squad modes as two faces of one tool, with every provider mark on the same bar.

The work

02 / 08

  1. Workbench. Projects, threads, and approvals on the left; engine, model, permission, and effort controls sit in the composer where the decision is made.

Cloaky agent workbench with projects, threads, and composer

The work

03 / 08

  1. Review. Read the diff, leave line feedback, ship when ready — the privacy tier stays pinned in the corner.

Cloaky review view with a diff ready to commit

The work

04 / 08

  1. Engines. Claude Code, OpenCode, Codex, Kimi Code, and Grok Build in one harness, each an explicit opt-in.

Cloaky engines view listing agent engines

The work

05 / 08

  1. Subagents. Every delegated task keeps its own inspectable feed, opened from a card in the thread.

Cloaky subagents running in parallel

The work

06 / 08

  1. Rewind. Take back the conversation, the files, or both — from any message you sent.

Cloaky rewind view restoring an earlier point in a thread

The work

07 / 08

  1. Bots. Named teammates with routines on schedules, on the same engine underneath — messaged like colleagues.

Cloaky bots view

The work

08 / 08

  1. Art direction. Painted scenes carry the brand’s calm, outdoors tone — this one depicts the local-model route.

Art direction. Painted scenes carry the brand’s calm, outdoors tone — this one depicts the local-model route.

What this case shows

A shipped product, described by what is built

Cloaky ships as desktop builds from cloaky.dev. This case shows the product surfaces and the positioning behind them; it does not claim user numbers or revenue.

What the work delivered

  1. 01

    Reframed the product from “private” to “pseudonymous” so every claim matches what each backend actually does.

  2. 02

    Designed a workbench where privacy tier and cost are visible per session, not buried in settings.

  3. 03

    Built the Base / Pro site and interactive product tour that explain the model without overstating it.

Source material

cloaky.dev

Next case study

02Know Your CodeStudio product / voice-first AI skill

How do you explain the code someone shipped without lecturing them?