prevolut.uk / Products / C.A.S.H.
Live in Chapper on iPhone & iPad

C.A.S.H. gives AI a real shell on iOS, without handing it the whole device.

C.A.S.H. is the controlled runtime behind Chapper's tool execution layer on iPhone and iPad. It lets AI inspect files, write results, automate structured local work, and verify what actually changed inside a bounded sandbox.

Shipping inside Chapper on iOS with sandboxed file access and clear path aliases.
Real shell patterns: pipes, redirects, chaining, variables, history.
Designed for AI-assisted work on iPhone and iPad, not arbitrary full-device execution.
Integrated with confirmation, file-change reporting, and chat-visible execution logs.
cash://sandbox · v1.0
Prevolut Runtime
cash ~ $ pwd /Documents/cash cash ~ $ echo "Latest News about Artemis" > $CASH/brief.txt
filesystem_changes created_files: brief.txt modified_files: none exit_code: 0
cash ~ $ cat $CASH/brief.txt Latest News about Artemis C.A.S.H. keeps the task local, inspectable, and recoverable.

A simple question: how do you let AI do real work without losing control?

Prevolut did not want a fake terminal, a toy file picker, or a vague "agent" abstraction. The goal was narrower and harder: real command-like power, strict boundaries, readable logs, and predictable recovery when something goes wrong.

C.A.S.H. began as the execution layer behind Chapper's more advanced workflows. The first problem was not "how do we run commands?" It was "how do we make command execution understandable, auditable, and safe enough for a mobile product?"

That pushed the design away from unrestricted system access and toward a dedicated runtime with explicit directories, explicit prompts, explicit confirmation, and explicit file-change reporting. The result is not a desktop shell clone. It is a product runtime built for structured AI work.

That distinction matters. A good shell for AI is not the same thing as a shell for a sysadmin. It has to explain itself, recover cleanly, and fit into an interface where users need to see exactly what happened.

Phase 1

Controlled file work

Start with sandboxed directories and deterministic file operations instead of broad device access.

Phase 2

Real shell semantics

Add chaining, redirects, variables, quoting, and command history so workflows stop feeling artificial.

Phase 3

AI execution discipline

Layer in confirmation, MCP integration, file verification, and visible tool banners inside the chat itself.

Now

In development, already real

C.A.S.H. is still evolving, but the core runtime already powers concrete AI-assisted shell and file workflows inside Chapper.

A shell runtime shaped around product constraints, not demo theatrics.

C.A.S.H. gives models practical execution tools while keeping the surface area understandable. The point is not maximum freedom. The point is reliable, explainable local work.

>

Real command structure

Commands support redirects, chaining, variables, quoting, history, and a usable subset of classic shell workflows, so generated actions stay practical instead of brittle.

#

Bounded file access

$CASH, $HOME, and related aliases point into a known sandbox. That gives AI enough room to create, inspect, and transform local files without pretending the whole device is fair game.

+

Verifiable outcomes

Command results include exit state and file-change metadata, so the model can confirm whether work actually happened instead of inventing success from thin air.

C.A.S.H. turns "AI can help" into "AI can complete the task."

When an assistant can search, gather, write, verify, and hand back a real file, the interaction changes. It stops being only conversational. It becomes operational.

Built for concrete local work

  • Create research notes, reports, snippets, and working files directly inside the Chapper sandbox.
  • Chain simple shell logic without forcing the user to leave the chat and manually repeat steps.
  • Expose what ran, what changed, and what still failed, instead of hiding execution behind a black box.
  • Keep the user in control with confirmation flows for sensitive command execution.
State

Live inside Chapper

The runtime is already shipping inside Chapper on iPhone and iPad and keeps getting stricter, clearer, and more capable.

Positioning

Not a toy terminal

This is not a generic novelty shell. It is an execution layer for product workflows where safety, traceability, and UI integration matter.

Design goal

Power with boundaries

Enough capability to get meaningful local work done. Enough structure to keep the experience legible and recoverable.

Platform

Built for iOS reality

Designed around mobile constraints, app sandboxing, and user-visible execution instead of pretending an iPhone should behave like a remote Linux box.

C.A.S.H. is one layer in a broader tool stack.

It works alongside MCP tools, memory, web search, and structured chat flows. Each layer has a role. C.A.S.H. is for the moment when the assistant needs to do grounded local work and leave artifacts behind.

Inside Chapper

C.A.S.H. integrates directly with the chat experience, tool banners, command confirmation, and model-facing tool prompts.

  • Visible execution logs inside the conversation.
  • Sticky state for active shell work and confirmations.
  • Tight coupling with AI tool orchestration instead of bolted-on terminal UI.

For future workflows

As Prevolut expands local AI tooling, C.A.S.H. becomes the base layer for repeatable file and command tasks that need more than plain chat.

  • Research-to-file pipelines.
  • Verification loops using filesystem changes.
  • Local task chains that stay inspectable from start to finish.

See where C.A.S.H. is headed.

C.A.S.H. is part of the broader Chapper roadmap. If you want to follow the product as it matures, start with Chapper itself and the public product pages around it.