OhMyBug

OhMyBug plugin — bughunter

[Discord]

We find the bugs your Claude Code missed. You pay only when your own Claude confirms real bugs: $4 per bug-finding review (however many bugs it holds), $12 for a deep hunt you choose ($24 on a wide change), minor findings free, first review free.

• Fast review · $4 · finds 73% of significant bugs · up to 15 minutes
• Deep review · $12 ($24 on a wide change) · finds 90% of significant bugs · up to 2 hours

Install (Claude Code)

claude plugin marketplace add OhMyX-AI/ohmybug-pluginclaude plugin install bughunter@ohmybug

That's the whole setup. On first use Claude Code opens a GitHub sign-in in your browser; your account (with a free first review) is created automatically. No keys to copy, nothing lands in config files.

Install (pi)

pi has no built-in MCP — add the adapter, then the server:

pi install npm:pi-mcp-adapternpx skills add OhMyX-AI/ohmybug-plugin --skill bughunter

Add to .mcp.json in your project root (or ~/.config/mcp/mcp.json):

{  "mcpServers": {    "ohmybug": { "url": "https://mcp.ohmybug.ai/mcp", "auth": "oauth" }  }}

Restart pi, run /mcp-auth ohmybug once — a GitHub page opens in the browser, your account with a free first review is created automatically. Then say "hunt bugs" or /skill:bughunter.

Prefer a raw key (CI)? Use "auth": "bearer", "bearerTokenEnv": "OHMYBUG_API_KEY" instead of "auth": "oauth".

Optional merge gate (pi has blocking tool hooks) — .pi/extensions/ohmybug-gate.ts:

import { isToolCallEventType } from "@earendil-works/pi-coding-agent";
pi.on("tool_call", async (event) => {  if (isToolCallEventType("bash", event) && /\bgh pr merge\b/.test(event.input.command)      && !process.env.SKIP_BUGHUNT) {    return { block: true, reason: "Final diff not hunted — run /skill:bughunter first (or SKIP_BUGHUNT=1)" };  }});

Install (Codex app / CLI)

The plugin package wires the skill, MCP server, and Codex review routing in one install. Codex uses the same GitHub OAuth flow as Claude Code:

codex plugin marketplace add OhMyX-AI/ohmybug-plugincodex plugin add bughunter@ohmybugcodex mcp login ohmybugcodex update

The last line brings Codex itself up to date, and it stays last: where Codex cannot tell how it was installed (for example, bundled inside another app) it exits with an error, and by then the plugin, the MCP server and the sign-in are already in place. Update that Codex the way it was installed.

Open /hooks once after installation, review and trust the OhMyBug hook definition, then start a new Codex thread. Codex skips non-managed plugin hooks until you trust their current definition; without that step the skill and MCP still install, but automatic review routing is not active. Then ask Codex to run the OhMyBug cloud hunt (or use the plugin's starter prompt). A local APPROVE from a code-review agent is not merge evidence.

Existing skill-only installs

If the plugin was installed with npx skills add, keep the skill and add the same MCP server explicitly:

npx skills add OhMyX-AI/ohmybug-plugin --skill bughuntercodex mcp add ohmybug --url https://mcp.ohmybug.ai/mcp

codex mcp add detects OAuth and opens the GitHub sign-in flow automatically.

For headless CI, use an operator-issued raw omb_ key — write to [email protected] or DM @ohmybug to get one:

export OHMYBUG_API_KEY=omb_...codex mcp add ohmybug --url https://mcp.ohmybug.ai/mcp \  --bearer-token-env-var OHMYBUG_API_KEY

(Claude Code users who prefer a key over OAuth can do the same with claude mcp add --transport http ohmybug https://mcp.ohmybug.ai/mcp --header "Authorization: Bearer $OHMYBUG_API_KEY" — an explicit Authorization header disables the OAuth flow.)

Questions → Discord.

First run: no pull request needed

Open your agent in the project and run /bughunter:review (or say "hunt bugs in my code"). It hunts whatever is there: a branch against main, uncommitted changes, the last commits, or – in a folder that is not a git repository at all – the files themselves as new code. It says which one it picked before anything is sent, and you can point it elsewhere ("hunt src/solver/", "the last 3 commits").

On a GitHub repository, /bughunter:scan hunts the code this repository changed in the last 30 days (/bughunter:scan 14 for two weeks). The server reads the history from GitHub and clones the repository on its side — nothing is uploaded from your machine — then runs a deep hunt (up to 150 minutes) over the window's changed code, and hands back one merged review that lists anything it did not read. Your agent verifies the findings like any hunt, and it is priced as a deep hunt: $12, or $24 on a wide change, only if a real bug is confirmed. It needs the OhMyBug GitHub App on that repository, runs once per repository per week, and is not a merge gate.

What it does

• The hunt is the review: run it as soon as a change is ready, instead of a local review. Before every gh pr merge, a gate is the safety net: it checks the final diff was hunted (skip once with SKIP_BUGHUNT=1; review fixes change the diff and re-arm the gate).
• /bughunter:review (or just ask to "hunt bugs") sends your diff + selected context to OhMyBug's cloud. You see the exact file manifest before upload.
• An orchestrated fleet of adversarial reviewers hunts on OhMyBug's models. Only findings that survive triage come back.
• Your own agent verifies each finding against your codebase and reports REAL / NOT_REAL / UNCLEAR. You see and approve the verdicts.
• Billing: one flat price per review with confirmed real bugs — one or ten, same price: $4 fast, $12 deep, $24 deep on a wide change (11+ folder chunks). Minor findings, review runs, false positives, unclear — $0.
• /bughunter:stats shows your record: bugs found, reviews run, balance.
• Before a diff or file contents leave your machine (submit_review, provide_files), a secret scan refuses a payload that carries a credential (vendor-prefixed tokens, private keys, key files) and says which file and line, with the value masked. A payload it cannot read is refused too. It lets a payload through unscanned in four cases only: submits that send no payload (repo + ref) have nothing to scan; without python3 it cannot run, and says so in a system message; it does not run on Windows; the out-of-band upload for very large diffs is not scanned. Set OHMYBUG_SECRET_SCAN=0 (in the env block of settings.json) to turn the scan off.

What runs where

Everything that blocks or refuses lives in shell hooks, so it behaves the same in every client. The Claude Code mod shows and offers; it never refuses. It draws a button only where you have to decide (the deep-hunt offer, which spends money); the agent's own watch picks up a finished hunt, so that band only says what happened.

A finished hunt someone already read or judged — in another chat, or in a way the mod did not see — gets no toast, band, resume or verdicts line here: the mod asks the status URL (a finished, unread hunt again every few minutes), and the server says whether the result was read or has verdicts. If the status URL cannot be reached, the band stays until the agent reads the hunt or your next turn starts.

The mod never starts a turn on its own unless you opt in. Two environment variables tune it (in the env block of settings.json, or managed settings for a team):

• OHMYBUG_AUTO_RESUME=1 — when a tracked hunt finishes, fails or asks for files and the session is idle, send that hunt's next step instead of showing the band. Off by default.
• OHMYBUG_STATUS=0 — keep the status line clear; the prompt line goes with it, so the agent is never told what the person cannot see.

The VS Code extension has no band above the prompt; there the toast and the status line carry it. Turn the mod off with /plugin like any built-in mod; the shell hooks stay.

MCP contract (server: https://mcp.ohmybug.ai/mcp)

Errors: payment_required (top up), pending_verdicts (resolve previous review's verdicts first).