Finds 6× more real bugs than Cursor on ReviewBench.
The bug hunter for your coding agent. Catches the bugs your agent missed, then your own agent confirms and fixes every real one.
69,179 bugs caught this week
9,389 today · 6,355 reviews this week · 154 species in the Bugarium
claude plugin marketplace add OhMyX-AI/ohmybug-plugin
claude plugin install bughunter@ohmybug
Fails CI on a medium+ bug. Other CI ↗
GitHub sign-in: your name only, no repo access.
Stop reviewing in circles. Close the loop.
Your agent reviews itself
- A long list: real bugs mixed with style notes and guesses.
- You check each claim against the code by hand.
- You re-run it, and a fresh list arrives.
Back to step 1.
Your agent + OhMyBug
- 1Your agent writes the diffas it does now
- 2OhMyBug hunts it in the cloudone deduplicated list; a second check drops weak claims
- 3Your agent verifies and fixesthe real bugs, in your repo
- Clean mergea hunt that finds nothing costs $0
Then the next diff, same loop.
What a hunt looks like
Illustrative session · sped upPick a speed. $0 if clean.
How to pick a mode ↗-
lite
$1
- Typical answer
- 85 s
- Known bugs found
- 61.5%
- Hard ones found
- 52.5%
For saving: a little more than snap finds, at half the price, and slower.
-
snap
$2
- Typical answer
- 34 s
- Known bugs found
- 57.7%
- Hard ones found
- 47.5%
For the loop, while your agent is still writing.
-
fast
$4
- Typical answer
- 1.9 min
- Known bugs found
- 76.5%
- Hard ones found
- 77.5%
The check before the merge.
-
deep
$12 $24 on a wide change
- Typical answer
- 4.6 min
- Known bugs found
- 86.1%
- Hard ones found
- 77.5%
Risky or wide changes: payments, security, data.
Measured on OMBench: 85 held-out changes with 742 known bugs, November 2026. Typical answer is the median time on our side. Hard ones are 40 bugs that earlier reviews missed. Lite, snap and fast are from later runs.
Bug recall on ReviewBench
We replayed 171 real bugs at the commit that introduced them
Score 138 of 171 80.7%
124 caught · 28 partial, each worth half · 19 missed
- django/django 59 fixes from 2026
- postgres/postgres 98 fixes from 2026
- public CVEs & OSS bugs 14
Sampled catches
- laravel/frameworkAuth token guard accepts array inputlived ~10 yearsCaught · deep hunt
- postgres/postgresHash bucket split leaks a buffer pinlived ~10 yearsCaught
- django/djangoHTML truncator denial of serviceCVE-2026-1285lived ~2 yearsCaught
- agno-agi/agnoCode execution via
eval()on model outputCVE-2026-35002lived ~7 monthsCaught - getsentry/sentry-javascriptAuth middleware bypass via the tunnel-route prefixGHSA-mchm-gjx6-7h55lived ~14 monthsCaught
- tokio-rs/tokioTimer data racelived 25 daysCaught
Bug species it catches
Browse Bugarium ↗- LOGIC

Fix on one copy only
Rule: Grep every writer.
- INPUT

Falsy id drops account 0
Rule: Check
id == null. - CACHE

Missed cache eviction
Rule: Share one key helper.
- LIMITS

Off-by-one quota guard
Rule: Test
cap&cap+1. - TXN

Webhook before DB commit
Rule: Outbox post-commit.
- UNITS

Seconds vs milliseconds
Rule: Suffix
_s/_ms. -
NEW SPECIES
Catch a bug nobody has named yet.
Become its discoverer — then name it and pick its mascot for the Bugarium.
Start hunting →
How it works
Technical note ↗Diff + context
Your agent submits the change; the hunt runs in an isolated cloud workspace.
Hunt + 2nd pass
Drops false alarms ($0).
Your agent verifies
Billed from $1 per review if REAL (1 or 10 bugs) · $0 if clean.
Give your agent the reviewer ranked
#1 on ReviewBench. as of November 11, 2026
Flat per review, never per bug — from $1 (lite), $2 (snap), $4 (fast) or $12 (deep). Clean and minor-only runs cost $0; false alarms you reject cost nothing (how billing works).