Your AI
Technical Engineer.
Build, fix, connect and keep it running. Every change arrives as a proposal you review, never as a surprise.
Early access is free while KeepRun is being built. Read-only by default. Sensitive actions always stay under your control.
Built with one of these? That’s exactly who this is for.
- Lovable
- Bolt
- Base44
Replit
Cursor
v0
Claude Code
- Codex
Supabase
Vercel
One engineer. Seven jobs.
Each card below says plainly whether we have already measured it end to end in our lab, or whether it is still being built. We will not blur that line.
-
Measured in lab
1 · Builds new features
Tell it what you want in ordinary words. It reads your code, writes a plan you approve first, then proposes the change for you to look at.
You approve the plan before anything is written.
-
Measured in lab
2 · Diagnoses and fixes bugs
It looks into the problem itself, tells you the cause, proposes a fix and tries it on a copy. If its first fix does not work, it reads why and tries again.
You get the cause, not just a patch.
-
In development
3 · Creates and changes APIs
It can write a completely new file and connect it to your app, which we have measured. Delivering a finished API that passes its own test is not signed off yet.
New code gets written for you, not just small edits.
-
In development
4 · Connects your systems
It can already read how your database is built without being allowed to change it. Connecting a real outside service from start to finish comes next.
It can look without being able to break anything.
-
In development
5 · Watches and spots incidents
Everything it does is already written down permanently. Watching your live app and warning you in time is being built on top of that record.
Nothing it does can be quietly undone later.
-
In development
6 · Security and maintenance
Code it writes is already tried out in a sealed box with no access to the internet and no special privileges. Looking after your app’s security day to day is not open yet.
Your code never runs on a machine that matters.
-
Measured in lab
7 · Human supervision and evidence
It never puts anything live by itself. Every run leaves you a change you can read, a result you can check and a cost down to the penny.
Nothing reaches your app without you saying yes.
What it has actually done
Three runs on our synthetic lab application, with the real model, real money and an independent criterion the model never saw. The numbers below are measured, not estimated.
-
It repaired a bug nobody told it how to fix
VerifiedGiven a failing behaviour and no solution, it investigated, proposed a patch, had it rejected by the independent check, read why, and corrected its own fix on the second attempt.
- 2 attempts
- $0.070868 total
- VERIFIED
- an empty patch was caught by a non-vacuity witness
-
It wrote a feature that did not exist
VerifiedA support request in plain prose, “answer what is this payment from a provider reference alone”, became a working read on the application. It was judged by a criterion that asserts values read independently from the tables.
- $0.106767 plan + code
- VERIFIED
- a stub that compiled and returned nothing was recorded as FAILED
-
It created a new module and wired it in
Partly verifiedIt wrote a 71-line module that did not exist at the pinned commit, imported it from the application file and declared the new read. The proposal applied, compiled, and left the project’s own thirteen tests green. The withheld criterion then failed on one of six checks, because it read a property name our own specification never stated. We are reporting that, not hiding it.
- +71 lines created
- 13/13 existing tests still green
- 5/6 criterion checks passed
- FEATURE NOT COMPLETE
These runs happened on a synthetic application we built to be measurable, not on a customer’s production system. KeepRun is not yet open to the public, and we would rather tell you exactly how far it has got than imply more.
See KeepRun in action
The first videos are coming soon. They will be recorded from the runs above as they actually happen, with no reconstructions and no mock-ups.
Want to be told when the first one is out? Join early access below.
You stay in control
This is not a promise about intentions. It is how the engine is built: there is no code path that applies a change without you.
- ReadKeepRun observes and investigates without modifying anything. Read access and write access are separate scopes, checked separately.
- ProposeChanges arrive as a diff computed against a pinned commit, verified on a throwaway copy inside a container. Your repository is never touched.
- ApproveApplying, committing, pushing and deploying are human steps. The agent has no privilege to do any of them.
- AuditEvery run records what it claimed, what was measured, why it failed and what it cost. The record is append-only and never rewritten.
Pricing
Not on sale yet. These tiers are the shape we are building towards, and the prices are provisional. We will not charge anyone until quotas and permissions are enforced on the server, not just drawn on this page.
Provisional. Nothing is billable today
-
Free$0
- Really test KeepRun, with limited quotas
- 1 project
- Read-only investigations
- Full evidence on every run
-
Starter$19.99 / month
- More credits per month
- Several projects
- Feature development and bug repair
- Attempt history and cost per attempt
-
ProTBD
- Higher quotas and more skills
- API creation and system integration
- Monitoring and incident detection
-
BusinessTBD
- Team seats and shared approvals
- Private infrastructure options
- Priority support
Getting in
KeepRun is in private beta. Rather than show you a sign-in box that cannot sign you in, here is exactly where things stand.
-
Not open yet
Already have access?
The dashboard already runs, and it shows real work: the plans, the changes it proposes and the test results. It is not on the internet yet, and there is no sign-in at all: accounts, sessions and separation between users are designed but not built. We will not open a door before the lock exists. If you were given access, we will send you the link directly.
-
This works
Want access?
Tell us what changed in your product. It goes to a real inbox and a person reads it. There is no automated sign-up behind this form, and it will tell you if it fails rather than pretending it worked.