Hire AI to do real work — in a room your expert never sees inside.
AI City rooms are sealed sandboxes where AI does the hands-on work. You invite a buyer (or a friend) by magic link. They watch — but the expert behind the room sees only what each execution mode allows: nothing, the AI's questions, or your code with names and strings scrambled.
No account required to hire. Pay by Stripe. Claim sessions to a dashboard afterwards.
Three execution modes
Trade depth for privacy
Pick how the work gets done when you open a room. Each mode has different privacy properties and different pricing. The buyer always sees the choice before paying.
Methodology
$A pre-built skill runs autonomously inside the sandbox. The expert isn't online during the room.
Privacy
Your code stays sealed. The expert never sees it — only the AI inside the sandbox does.
Best for: Code review, audits, reviews where the expert's pre-built process is the value.
Consult
$$AI does the hands-on work; the expert is on call to answer the AI's questions in real time.
Privacy
Your code stays in the sandbox. The expert sees the AI's questions, never the code itself.
Best for: Tricky problems where the AI needs human judgement to navigate decisions.
Glass-Box
$$$The expert drives directly via MCP — but on a redacted view. Variable names, identifiers, and string literals are scrambled before they see anything.
Privacy
Strong privacy, not airtight. The expert sees structure but not naming, comments, or your repo URL.
Best for: Deep refactors and architecture work where you need a human at the keyboard.
Privacy is the product
Your code never leaves a sealed sandbox
The moat is what your expert doesn't see. AI City makes the privacy guarantee architectural — not promised in a contract.
Sealed sandbox
Every room runs in an isolated microVM. Network blocked, files read-only, environment destroyed when the room closes.
Mode-aware visibility
What your expert sees depends on which mode you pick. Methodology: nothing. Consult: questions only. Glass-Box: redacted code only.
Audit log
Every action inside the room is recorded — files read, edits proposed, tests run. You can replay it after the room closes.
Optional escrow
Hold the work fee with AI City via Stripe Connect. Released to your seller on accept, refunded on reject. 4-5% fee — only when used.
Two ways in
You can hire from AI City. You can also sell on it. Both share the same room primitive.
For buyers
Open a room. Watch the AI work via the magic link. Accept or reject — full refund on reject.
- /hire — one page progressively asks buyer or seller, then which flow.
- Auto-agent — an AI runs in the sandbox, you watch via the magic link.
- Invite an expert — a human drives in Glass-Box mode via Claude CLI on a redacted view.
For sellers
Sign up. Set up your seller profile. Open rooms for buyers, or list yourself in the public marketplace.
- /hire?as=seller— open a room for a buyer you're already talking to.
- Public marketplace — toggle a profile on for buyers to discover you.
- Stripe Connect — escrow payouts arrive in your own bank account, not a platform balance.
$ claude mcp add ai-city npx -y @ai-city/mcp@1.1.0 \
--env AICITY_SESSION_ID=ses_… --env AICITY_CLAIM_TOKEN=cl_…
✓ Driver tools loaded (claim_room, list_files, propose_edit, run_tests…)
you: claim my room and refactor the auth module
↳ Reading redacted files…
↳ Proposing rename: fn_a17 → validateSession
✓ Tests passing — delivering room
Drive a room from your CLI
Glass-Box
via MCP
Buyers can hire a human friend to drive a Glass-Box room directly from their own Claude CLI. The MCP package exposes 9 driver tools — read redacted files, propose AST renames, run tests, deliver. The buyer's un-redacted code stays in the sandbox; the friend never sees it.
Common questions
What is a 'room'?
A room is one task, one magic-link session, one audit log. A sealed sandbox spins up, work happens inside, the buyer watches via the Observer link. When the work is delivered, the buyer accepts (releases payment) or rejects (full refund). Then the sandbox is destroyed.
Who runs the AI?
It depends on the execution mode. In Methodology and Consult, the platform runs an LLM inside the sandbox — your code never leaves and the expert never touches it. In Glass-Box, the expert drives via MCP from their own Claude CLI, but only sees a redacted view (renamed identifiers, stripped comments). Three different privacy properties — pick the one that fits.
Can the expert see my code?
No, never the original. In Methodology mode, only the platform's AI inside the sandbox sees your code. In Consult mode, the expert sees the AI's questions, not the code. In Glass-Box mode, the expert sees a redacted twin — variable names and string literals scrambled before they see anything. The un-redaction happens inside the sandbox at delivery.
How much does this cost?
Every room has an upfront room fee (sandbox + duration + LLM cap, charged at booking via Stripe). Sellers set their own work price. There's no flat platform fee on top. If the buyer enables escrow, AI City takes a small fee on the work fee (5% Free / 4% Pro / 4% Top). See /pricing for the breakdown.
What happens if I'm not satisfied?
Two paths. Reject the work before accepting and you get a full refund automatically — no admin review. After acceptance, you can file a dispute via the observer console; an admin reviews the audit log and decides on a full, partial, or no refund.
Do I need an account?
No, not to start. Open a room at /hire and pick auto-agent (an AI does the work) or invite an expert (Glass-Box: a human via Claude CLI on a redacted view of your code). Pay by Stripe Checkout. Your magic link arrives by email; claim the session into a permanent account afterwards if you want it on your dashboard.
Open a room. Send the magic link.
No subscription to start. Pay per room. Optional escrow. Refund on reject.