For AI-generated repo work you do not want to run blind

Approve AI-generated repo work before anything runs.

ApproveOnce is a local approval-and-execution layer for AI coding work. It turns “the model wants to do something” into one sealed plan you inspect first, approve once, run locally in isolation, and verify afterward.

Buy ApproveOnce Inspect proof

Refactors

See the repo change before it lands

When AI wants to touch real files across the repo and you want the exact plan first.

Local commands

No hidden shell steps

When execution matters and you do not want surprise installs, extra scripts, or vague runtime behavior.

Trust-sensitive work

A stricter boundary than careful prompting

When silent repo mutation is the risk and personal discipline alone is not a strong enough operating rule.

See the loop immediately

Plan. Approve once. Safe local run. Proof.

One real founder QA pass, not a diagram of a future product.

01 Plan

Sealed review plan

The model drafts one local plan before any write or command runs.

- append README.md
- shell: none
- network: none

02 Approval

Approve once

Approval binds to one exact plan hash and one workspace snapshot.

plan hash
89dc28f1be02...

git snapshot
73be2f15c01b...

03 Safe run

Local run in isolation

The change executes in safe mode before anything is applied back to the workspace.

status: succeeded
changes: 1
applied_to_workspace: true

04 Proof

Proof left on disk

The run leaves durable local evidence instead of asking you to trust memory alone.

approval.json
diff.patch
manifest.json
result.md

Plan

Draft one sealed local plan before any write or command runs.

Approve once

Inspect exact changes, commands, and permissions, then approve only that plan.

Safe local run

Execute locally in isolation instead of inside a vague remote agent loop.

Proof

Keep changed paths, logs, and the durable run record afterward.

What it is actually for

Not more autonomy. More explicit control.

ApproveOnce is for the moment an AI coding tool wants to do real work in your repo, but you do not want to trust it with silent execution. The wedge is review-before-mutation.

Instead of Prompt the model, hope it behaves, and reconstruct what happened after the fact.

You get One sealed plan, one explicit approval, one isolated local run, and one proof trail you can inspect afterward.

Best fit Founders and power users already using AI coding tools who do not want repo mutation to depend on caution alone.