Where does Builder run? · AI Workspace

What Builder mode is

Today, staff use Assist mode — the chat workspace that's already live. Builder mode is the graduate version: each person gets their own private workspace where an AI can actually do the work, not just talk about it — write code, run it, build and ship real things. To make that work, each person needs a small private computer in the cloud for the AI to work in. The question was simply where those live.

A real workspace

A proper computer the AI can use — run commands, keep files, install what it needs.

Private to each person

One person's work is walled off from everyone else's.

Connected

Can reach the AI, pull in code and tools, and publish what it builds.

On when needed, off when not

Spins up when someone works, shuts down when they're done — so it's not costing money idle.

The plan — Mosaic now, our own box later

Two stages. The first adds no new costs and uses only things we already have. The second is an upgrade for when we outgrow it or want full control.

Now · Phase 3

Run it on Mosaic

The platform that already hosts the AI Workspace app.

Builder runs inside Mosaic, right alongside the app. Nothing new to sign up for, no new vendor to manage. For our own staff — people we trust — this is plenty secure to start.

✓ No new bills — just Anthropic for the AI + Mosaic, both of which we already pay for

Later

Move to our own box

A dedicated machine that's ours.

When we outgrow shared hosting — or want full control and the strongest security walls — we move Builder onto our own dedicated box. Same software, just our own hardware. This is also what an outside client company would expect if we ever sell this.

↗ An upgrade we choose when it's worth it — not needed to start

A couple of things to confirm first

Small, honest caveats before we build — nothing that changes the plan.

Check with Mosaic: Mosaic hosts long-running apps today, not the "one fresh private computer per person" setup Builder needs. We need to confirm Mosaic can give us that — or a dedicated slice — and if not, we build it on top of what they have. This is the main thing to nail down.
Set a sensible limit per person: so a runaway task can't quietly rack up time. This plugs into the budget controls we already designed.
When "our own box" comes: decide whose box (ours, or hosted-but-dedicated) and where it lives — a later conversation, not a now one.

Other options we looked at

For the record — there are paid outside services built exactly for this. We're not using them, because they'd mean a new bill and a new vendor, and Mosaic covers us. Worth revisiting only if a future client demands their data stay fully on their own systems.

See the outside services we compared

These are specialist "sandbox" providers — companies whose whole job is renting private cloud computers for AI agents. All solid; all an extra monthly cost we don't need to take on now.

Service	Security walls	Rough cost	Why not now
E2B	Strong (own kernel)	~$150/mo + usage	New paid vendor
Vercel Sandbox	Strong (own kernel)	Pay-per-use	New paid vendor; can't ever be our own box
Northflank	Strong; can be our own box	Usage-based	Overkill until the "own box" stage
Modal / Daytona	Medium by default	Pay-per-use	New paid vendor; weaker default walls
Anthropic-hosted	Strong (Anthropic runs it)	Anthropic only	Still in beta; we'd rather keep it on Mosaic

Costs are rough vendor estimates, Nov 2026 — directional only. The full technical write-up lives with the AI Workspace planning docs.