The compliant runtime
Every AI-built app, inside the boundary.
Overnight is the compliant runtime for AI-generated software. Build with any model, IDE, or code generator. Security approves Overnight once, and every app that runs on it is safe to hold real company and customer data.
Deny by default/Per-tenant isolation/Signed provenance/Exportable audit trail/Any model/Real data/One approval/
No human actually reviewed this code.
Agents are shipping software on their own, against real customer data, faster than any review process can keep up.
The Overnight SDK decides what your model can do.
Your model never sees a raw key. It gets the capabilities you granted, and anything past that list never compiles.
Everything runs inside one boundary.
Every call the app makes takes one mediated path, and it gets logged on the way through. The boundary holds even for code that never touched the SDK.
What runs is exactly what you approved.
Every app is signed before it runs. Drift from the version you approved, and it simply does not start.
Run it inside the boundary.
Approve the boundary once and everything your team builds afterward arrives already inside it.
0 Security approval, ever
0 Raw keys your model sees
0% Of executions logged
0 Postures inherited: SOC 2, HIPAA, GDPR
Your team has been working around its tools for years.
Every workplace runs on software someone else designed, priced by the seat, and renewed whether or not it fits. Teams spend years bending their work to match it. Now a model can write the tool they actually wanted in an afternoon, which leaves exactly one thing missing: somewhere that code can safely run.
What changes on the invoice is the second-order effect. Every tool your team builds is one it stops renting, and what replaces the licence is runtime you are metered on.
stays you rebuild conforming to it 1,000 people · 87 apps
athe stack$6.0M of licences
bthe swap2 hrs/person/wk × $65
87apps in the stack
$12.0Ma year, all in
2 hrsa week, per person
53 / 87apps rebuilt
$4.1Msaved a year
−34%of the true cost
- 01 what you pay for today
A license for a fit you never got
Seat pricing bills you for access, then asks your team to work someone else’s way. The one workflow you needed is always the feature that never ships.
- 02 what changed
Software shaped like your company
The tool that matches your process has always been buildable. It was just never worth the quarter it would take. When a team can build it in a session, that math stops holding.
- 03 what is still missing
Somewhere to actually run it
Anyone can produce an app now. Nobody hands you a place it can touch customer data without becoming your next incident. That is the half Overnight builds.
Two products. One boundary.
The SDK decides what your model is allowed to write. The Runtime decides what is allowed to run. Approve that boundary once, and everything your team builds afterward arrives already inside it.
It starts as a conversation. You describe the tool, answer a few questions, and the answers become the list of what this app may do — alongside the far longer list of what it may not. Nothing your team builds afterward gets to argue with that list.
✳ start · overnight curate
1
step one · describe
What should this app do?
start curating
describe A tool for our support team to look up a… edit
2
questions
step two · questions
A few questions, support included
Got it. Your answers decide which capabilities land in your app.
Which systems does it read from?
Customer records Order history Shipping
Can an agent move money?
No, read only Refunds, up to a limit Any adjustment
drafting the manifest from 3 answers…
questions 3 answers · support edit
3
proposal
step three · app spec
What it may do in the runtime
▢ refund-desk
runtime manifest ·
24
allowed
allow
allow
allow
allow
block
proposal refund-desk · 4 capabilities edit
4
generate
last step
Hand it to your AI IDE
Claude Code Cursor Codex
CLAUDE.md # refund-desk — build brief Use @overnight/sdk. The runtime allows customers.read, orders.read, refunds.write, audit.write. Nothing else compiles. Refunds are capped at the limit set in the manifest; the runtime rejects the rest. Never read a credential — ask for a capability and the runtime hands one over.
generate app arrives inside the boundary security approved
01 the model-facing layer
Overnight SDK [O.N.S.D.K]
The SDK loads into your model as context and makes the compliant path the easiest one to take. Your model works from granted capabilities instead of raw credentials, so anything outside the approved manifest never compiles. You find the mistakes on a laptop instead of in production.
Explore the SDK → 02 the enforcement layer
Overnight Runtime [O.N.R.T]
The Runtime enforces the boundary around everything that executes. Each customer gets an isolated tenant, every call travels one mediated path, and each run is signed and written to an audit trail you can export. An app that never touched the SDK still cannot step outside the rules you set.
Explore the Runtime →Start with yourself.
Someone on your team needs a tool on Monday. They build it with the AI they already use, and it is running on real data by lunch, in a tenant that was compliant before they opened the editor. Nobody had to book a vendor call, run a pilot, or wait a quarter for procurement.
The second app is the one that proves the point. It lands in the same tenant, under the same identity and the same audit trail, and nobody has to review any of that twice.
tenantacme.onrt.ai approved once
RD refund-deskby support deploying ready
Look up a customer, refund an order, leave a trail.
- Waiting for environment
- Applying configuration
- Starting up
- Routing
open refund-desk ↗
IR invoice-reconby finance ready
Match statements against invoices every morning.
inherited the boundary · no second security review
open invoice-recon ↗
OT onboarding-trackerby people ops ready
Every new hire's first fortnight, in one place.
published to 3 teams · the boundary travelled with it
open onboarding-tracker ↗
1 apps 1 tenant 1 approval the same identity, isolation and audit trail under all of them
- 01 land
Build for yourself
The first app is the one your own team has been asking for. It runs in a tenant that is compliant from the first minute, so the value lands in the first session.
- 02 expand
Inherit the boundary
The second app arrives inside the same approved runtime, with the same identity, isolation, and logging. Security reviews one thing, once, no matter how many apps follow.
- 03 graduate
Publish to others
When something works, hand it to another team. The boundary travels with it, so the people receiving it skip the review you already passed.
What ran?/Who approved it?/What did it touch?/Under which identity?/
Something has to be accountable for what ran.
A model wrote it, nobody reviewed it, and it is already on production data. When the auditor asks what ran and who approved it, you have no answer. Overnight turns that liability into a boundary your auditors can sign off on.
The record is written while the app runs, so nothing has to be reconstructed afterwards, and it holds the refusals as well as the calls that went through.
apprefund-desk identityj.okafor · support events0
time call against decision
09:41:02 customers.read customer 8842 allow
09:41:02 orders.read order A-40913 allow
09:41:03 refunds.write order A-40913 · $40.00 allow
09:41:03 smtp.send not on the manifest denied
09:41:04 audit.write sealed · signed by tenant key allow
✕ denied The app asked to email the customer. Nobody granted that, so the boundary refused it — and the refusal is in the record too.
✕ ungoverned · audit exposure
- ✕AI-generated code touches customer data with no owner and no review.
- ✕Nothing recorded what ran, so you cannot prove who signed off on it.
- ✕Every shadow app is a SOC 2 finding waiting to be written up.
- ✕Compliance review stalls the launch, or the launch skips the review.
- ✕Each new tool restarts the whole security conversation from zero.
✓ compliant · audit-ready
- ✓One approved boundary owns identity, data, isolation, and egress.
- ✓Every call is mediated, isolated, and logged into an exportable trail.
- ✓The controls SOC 2, HIPAA, and GDPR ask about, carried by the boundary.
- ✓Overnight signs on as your runtime subprocessor of record.
- ✓Anything you did not explicitly allow is denied by default.
Approve the boundary once.
No seats to count and no license to renew before you know it fits. Start free, pay for the runtime you actually use as you grow, and move the whole thing into your own cloud whenever you want it there.
00 free
$0
Build one tool and run it inside the shared, compliant runtime. Nothing to sign.
- ✓Run one app inside Overnight.
- ✓Get help from the community.
01 premium · recommended
Usage
The whole platform, fully managed, for a company that builds its own software.
- ✓Build and run as many apps as you need.
- ✓Share the compliant runtime with other teams.
- ✓Use every premium feature we offer.
- ✓Pay only for the runtime you actually use.
- ✓Reach our team with priority support.
02 enterprise
Custom
Overnight in your own cloud, on your own terms, in a tenant that is entirely yours.
- ✓Everything included in Premium, and more.
- ✓Get a fully isolated tenant of your own.
- ✓Run it self-hosted in your own cloud or VPC.
- ✓Connect SSO and SCIM, and export full audit logs.
We are calibrating pricing with our design partners ahead of the public launch.
See full pricing details →Everyone else solves one piece. We solve the whole thing.
Code generators write the software, app builders assemble it, and cloud platforms host it. None of them takes responsibility for what that software does to your production data once it is running. That is the job Overnight took.
| What you actually need | Code generatorsCursor, Lovable, v0 | App buildersRetool, Replit | Cloud platformsVercel, AWS | Overnight |
|---|---|---|---|---|
| It runs the software it produces directly against your real production data. | ✕ | – | ✓ | ✓ |
| It gives every customer a fully isolated tenant instead of a shared pool. | ✕ | ✕ | – | ✓ |
| It signs on as your subprocessor of record, so compliance actually transfers to it. | ✕ | – | ✕ | ✓ |
| It keeps a complete, auditable record of everything that ran and who approved it. | ✕ | – | – | ✓ |
| You pass security review once, and every future app inherits that approval. | ✕ | ✕ | ✕ | ✓ |
| It stays responsible for the software once it ships, for as long as it runs. | ✕ | ✓ | ✓ | ✓ |
✓ Built in – Partial, or something you build yourself ✕ Not offered
See the full comparison → Questions security teams ask.
The blockers a platform engineer raises before AI-generated software touches real data, answered.
01 What is a compliant runtime for AI-generated code?
An execution environment where AI- or agent-written code runs with isolation, a full audit trail, and policy enforcement applied by default. The code is safe to run on real data and can pass a security review without per-tool engineering. That is what Overnight provides.
02 How do I run AI-generated code safely in production?
Run it somewhere that isolates it, logs every action attributably, and enforces what it is allowed to do before it touches real data. Overnight does all three by default, so you are not assembling that yourself for every new workload.
03 How is this different from an AI agent sandbox like E2B, Daytona, or Modal?
Sandboxes optimize for fast, isolated code execution during agent iteration and evals. Overnight is a production runtime. It adds the governance layer a security team needs before AI-built software runs on real data: a per-execution audit trail, runtime-enforced policy, and the controls SOC 2 asks about, enforced by the runtime.
04 Do we have to change how our engineers build?
No. Point Cursor, Claude, or your own in-house agents at the SDK and keep the workflow your team already likes. The SDK shapes what those tools are allowed to generate, and the Runtime enforces the same boundary again when it executes.
05 Is our data in a shared environment with other customers?
No. Every customer runs in a dedicated, isolated tenant instead of a shared pool, so one customer’s app can never see or reach another customer’s data, even by accident. Enterprise customers can run the whole thing self-hosted in their own cloud or VPC.
06 How does Overnight audit AI agent code execution?
Every execution is logged and attributable: what ran, what it accessed, and under which identity. The trail is written as it happens, so when an auditor or a customer asks what touched their data, the answer is one export away.
07 Is Overnight SOC 2 compliant?
Overnight is built to the controls those frameworks ask about: per-tenant isolation, an attributable audit trail, deny-by-default policy, and signed provenance, all enforced by the runtime itself. Overnight also signs on as your runtime subprocessor of record. Where any certification currently stands is a question we answer directly, in writing, whenever you ask it. Read how compliance and audits work →
Let’s build it
The hard part is where it runs.
Anyone can generate software now, which makes the generating part cheap. The scarce thing is a boundary your company trusts enough to point at real data.