Dungbeetle

Test anything, anywhere, instantly.

The regression safety net for AI agents and the humans they work for — web, desktop, terminal, anything. Zero adoption cost, runs anywhere.

Start for free Quick start Examples

agent loop · build → baseline → approve

# You → agent: build a hello-world index.html

🤖 wrote index.html

# Install Dungbeetle and capture the first baseline

$ npm i -g dungbeetle

$ dungbeetle update

📸 baseline saved · web:index

# Nothing changed yet — the check passes

$ dungbeetle test

✅ PASS 1/1 · no changes ⏱ 0.3s

# You → agent: change the heading to "Hello, Dungbeetle"

🤖 edited index.html

# Dungbeetle catches the edit as a semantic diff

$ dungbeetle test

❌ FAIL web:index — 1 diff

- <h1>Hello, world</h1>

+ <h1>Hello, Dungbeetle</h1>

# Looks intended — you review it and approve the new baseline

$ dungbeetle update

✅ approved · baseline updated

Watch a swarm work one branch

Four coding agents — Claude Code, Cursor, Codex, and Gemini CLI — build a docs site on the same branch at once. Every change is gated by a real visual test: the passes push straight through, and the one that quietly breaks the hero's contrast lands in your review queue instead of on main.

dungbeetle.dev · vibe cockpit

See how the agent loop works, step by step →

Features

🤖 Built for AI agents

Claude Code, Cursor, Codex, and Gemini CLI connect over MCP with scoped, human-approved tokens. Agents list runs, read diffs, and triage their own regressions.

🎯 Semantic diffs, not pixel noise

Output is normalized to kind-tagged JSON, so agents and humans review what changed — not a wall of pixels. Baselines survive cosmetic churn.

🔐 You own the account. Always.

Agents never hold accounts — they hold tokens you approve in the browser, scoped to one repository, attributed by name, revocable in one click.

🖥️ One tool, every surface

Web, desktop, terminal, API, performance, checks, and game walkthroughs — all captured the same way and reviewed like code.

⚡ Zero setup, runs anywhere

No browser needed for the default web fetch, and no external service for the CLI. Drops straight into CI — or your agent's sandbox.

✅ Humans approve

A central review/approve UI with flakiness analytics. Agents propose; you decide — every agent action lands in the audit trail under its own name.

Set up your repository

Install the CLI, describe what to capture, and save your first baselines. From then on, dungbeetle test catches any regression — on your machine, in CI, or in your agent's loop.

npm i -g dungbeetle

dungbeetle init      # 1. scaffold dungbeetle.config.json, then add your capture targets
dungbeetle update    # 2. capture those targets and save them as baselines
dungbeetle test      # 3. compare against baselines — exits non-zero on any diff

Give your agent a safety net

Repository set up? Connect your coding agent next. It never gets your password or your account — it gets a scoped token that you approve in the browser and can revoke any time.

dungbeetle login     # 1. prints a one-time code — approve it in your browser

# your agent now works on a scoped, revocable token:
dungbeetle test      # 2. it checks its own changes against your baselines
dungbeetle push      # 3. runs land in the review UI for your one-click approval

Works from Claude Code, Cursor, OpenAI Codex, and Gemini CLI over MCP — or from any CI. See the MCP guide, or everything an agent can read — llms.txt, markdown docs, and device-flow auth.

Cloud — start for free

When you'd rather not commit baselines, the hosted cloud stores runs centrally with a review UI and analytics. Free tier: 1 repository, 2 GB storage, 5,000 snapshots, 14-day retention.

☁️ Central storage

Runs and baselines stored centrally — nothing to commit to your repo.

👥 Team & history

Unlimited seats, with retention keeping runs as long as your plan allows.

📦 Artifacts & agents

Content-addressed screenshot storage, an MCP server, and scoped agent tokens you approve and revoke from the dashboard.

Start for free