It ships inside the assistant
It’s part of every AI assistant we build. See the assistant page.
Platform
Let AI run code, safely.
Let your AI assistant run calculations, check files and fill templates, inside strict limits you set.
Included with AI assistants and agents.
The shift
When a reply isn’t enough, your assistant can do the work.
Copying numbers out of a chat to run them by hand is slow. Giving AI free rein is risky.
We give it one job, specific files and a time limit. If it hits a limit, it stops and tells you.
The outcome
What you get
A task, not a blank console
The job is specific: this calculation, this check, this draft. A request to do something else is refused.
Bounds you can point at
Time, memory, and which files it may read or write are set before it runs. Those bounds are part of the plan.
A result you can read
The output comes back to the assistant with what it used. You can run it again or throw it out.
Stops safely
If anything looks off, it stops and explains why.
Where it earns its keep
Use cases
A calculation from rows you trust
The assistant takes numbers already in the file and returns the total the rule describes.
A check before something is sent
A file is tested against the rule you wrote: required fields, a total that matches, a name that exists. Failures go to a person with the reason.
A draft from a template
The task fills a template already on the list. It does not invent a new template, and it does not send the result. Sending stays a separate step.
The operating model
How it works
From first conversation to a system that runs itself.
Plan
We map how the work gets done today, including the calls only a person should make, and agree on it before anything is built.
Build
Built on Cloudflare and connected to the tools you already use, so nothing has to move.
Review
Tested on your real work from last week before any customer sees it.
Launch and run
We launch with close monitoring, keep it running, and you can change the rules at any time. It’s yours to keep.
The decision
How it compares
Letting a model loose on a laptop, hiring someone to run the sheet, and a fenced task are different risks.
| Off the shelf | Hire in-house | Abracadabra | |
|---|---|---|---|
| What it can touch | Whatever the plugin or the laptop can touch. | Whatever that person can open. | The files and the actions on the list, for the time that was set. |
| When it fails | The miss may show up later. | They notice if they are looking. | The run stops and the reason is in the log. |
| Who sends the result | Sometimes the tool does, if the toggle was on. | The person, when they are trusted to. | A person, unless sending is explicitly on the list. |
| When you don’t need us | The assistant only needs to talk, or the task is one formula in a sheet a person already opens. A code runner does not belong behind a conversation that never needed one. | A person running a weekly sheet by hand is fine when the sheet is small and the cost of a miss is low. | This stays out of the assistant when the job is conversation only. The sandbox earns its place when a task is repeated and a wrong write would be expensive. |
Named work
Proof
Running in production today.
Pricing
Priced to the project. Tell us what you need and we’ll quote it.
Questions, answered
FAQ
faq-title
Is this a server to rent and forget?
No. It is a task the assistant runs, inside a sandbox, on Cloudflare. An open machine and a password are not part of the delivery.
Can it download anything it wants?
No. It only uses tools you’ve approved.
Can it reach production records?
It only sees the data you allow, nothing more.
What if the model writes bad code?
The task still runs inside the sandbox, and the result is not sent until the review you set. A crash is a log line, not a change to the record.
Do you have to read the code?
You read the task in plain language: what it may open, what it may write, and when it must stop. We maintain the code.
Can this be bought without an assistant?
It comes as part of an AI assistant.
Start a project
Tell us about a task you still do by hand after a chat, and we’ll show you how your assistant could do it safely.