Each review in its own sandbox, destroyed when it ends.
A review copies your diff and files into a sandbox of its own, with no way out but the model and our API. When the review ends, the sandbox is destroyed with the copy in it. Copies outside the sandbox and the records we keep are covered in our privacy policy.
01 · Review time
Typical review time on our benchmark.
Each review runs in its own isolated VM sandbox, and the VM is destroyed when the review ends (terms).
snaptypically 34 s
litetypically 85 s
fasttypically 1.9 min
deeptypically 4.6 min
10 s1 min10 min1 h
Median time of one review per mode on our OMBench benchmark (November 2026), on a log scale. Each attempt is capped at 6.5 hours (terms).
Copied in. The diff and files go into a fresh sandbox.
Reviewed. The sandbox can reach the model and our API, nothing else.
Destroyed. The review ends, the sandbox and the copy inside it are deleted.
02 · How much of your repository leaves
How much a partial copy sent on our benchmark.
When we cannot read your code on GitHub, our plugin by default uploads a partial copy if the review can work from one: the changed files, the files that use the names you changed, and the build manifests around them.
1.9% median share on our benchmark0.2% median share, larger half of its repositories
Each square is 0.1% of the repository's size. Measured on fast reviews only, with the benchmark's copy of the file-picking rule, on 54 changes: the median share was 1.9% of the bytes git tracks, and a median change sent 25 files and 38 kB. In the larger half of those repositories (median 14.6 MB), the median share was 0.2%. A deep review reads the whole repository, from GitHub or from a copy you upload.
03 · What uploads leave out
Uploaded copies leave out secret-named files.
Every copy your agent uploads, partial or whole, leaves these out, in any letter case:
ignored files not committed
submodules
.env*
*.pem
*.key
id_rsa*
id_ed25519*
.npmrc
.netrc
.pgpass
.git-credentials
.ssh/
.aws/
.docker/
kubeconfig
*.tfstate*
*.tfvars*
*secret*
*credential*
The filter goes by file name: a secret in a file under another name is not caught by it. It applies to uploaded copies: a file you committed is read like any other when we review from GitHub, and a file a reviewer asks for during the review is sent as it is if your agent sends it.
The details
A sandbox of its own
Each review runs in its own throwaway virtual machine. Its network is deny-all except the model endpoint and our API, and the network layer enforces it. If the lockdown cannot be applied, the review does not start.
The sandbox is destroyed when the review ends.
No training on your code
We never train models on your code. The model providers process it under their own API terms, which do not permit training on customer content; those are their commitments, set out in our privacy policy.
Before some reviews, a classifier picks likely bug types. It receives the changed file paths and line counts, never lines of code.
Kept after the review
Records such as run transcripts, findings and the review record are kept after the review ends, and findings and reports can quote code. The full list, and how long each is kept, is in our privacy policy. To delete your account, email ohmybug@ohmyx.ai; the privacy policy says what must stay.