What Gasoline Captures

Gasoline is an open-source MCP server that passively observes your browser, analyzes performance, and generates code — everything an AI coding assistant needs to diagnose and fix issues autonomously. Zero dependencies. Localhost only.

Console Logs

All console API calls captured with full argument serialization:

Method	Level
`console.error()`	error
`console.warn()`	warn
`console.log()`	info
`console.info()`	info
`console.debug()`	debug

Objects, arrays, and Error instances are fully serialized — not just [Object object].

Network Errors

Failed API calls (4xx and 5xx) captured with:

HTTP method and URL
Response status code
Response body (the actual error payload)
Request duration (ms)

Your AI sees why the API call failed, not just that it failed.

Exceptions

Uncaught errors and unhandled promise rejections with:

Error message
Full stack trace
Source file, line, and column
Source map resolution (minified → original)

WebSocket Events

Full details →

Connection lifecycle (open, close, error)
Message payloads (sent and received)
Adaptive sampling for high-frequency streams
Per-connection rates and schemas

Network Bodies

Full details →

Request payloads (POST/PUT/PATCH)
Response payloads for debugging
On-demand — doesn’t record everything

Full interaction recording with 6 action types:

Action	What’s Captured
click	Element selector, URL, timestamp
input	Field selector, value (passwords redacted)
keypress	Key name, target element
navigate	Destination URL
select	Selected option, target element
scroll	Scroll position (throttled)

Smart selector priority: data-testid > role > aria-label > text > id > CSS path

Screenshots

Auto-captured on error as JPEG:

Rate limited: 5s between captures, 10/session max
JPEG quality 80%
Triggered by exceptions and console errors

AI Context Enrichment

Errors enriched with framework-aware context:

Component ancestry (React, Vue, Svelte)
Relevant app state snapshots
Custom annotations via window.__gasoline.annotate()

Web Vitals

Core Web Vitals captured and assessed against Google thresholds:

Metric	Good	Poor
FCP (First Contentful Paint)	< 1.8s	≥ 3.0s
LCP (Largest Contentful Paint)	< 2.5s	≥ 4.0s
CLS (Cumulative Layout Shift)	< 0.1	≥ 0.25
INP (Interaction to Next Paint)	< 200ms	≥ 500ms

Plus: Load Time, TTFB, DomContentLoaded, DomInteractive, request count, transfer size, long tasks, and total blocking time.

Performance Regression Detection

Automatic baseline computation and regression alerting:

Baselines computed from rolling averages (weighted 80/20 after 5 samples)
Regression thresholds: >50% load time, >50% LCP, >100% transfer size
Causal diffing — identifies which resource changes caused the regression
Actionable recommendations generated automatically
Push-based alerts when regressions are detected

API Schema Inference

Auto-discovers your API structure from captured network traffic:

Groups requests by endpoint pattern (normalizes dynamic segments)
Infers request/response shapes from observed payloads
Exports as OpenAPI stub or compact gasoline format
Configurable minimum observation count before including an endpoint

Session Checkpoints

Save and compare browser state over time:

Named checkpoints — save state at meaningful moments (up to 20)
Auto-checkpoint — implicit baseline for continuous monitoring
Timestamp queries — check state at any ISO 8601 time
Compressed diffs — only changed data returned (console, network, WebSocket, actions)
Deduplication and fingerprinting for efficient comparisons

Noise Filtering

Keep your AI focused on real issues:

Auto-detect — identifies repetitive patterns and suggests dismissal
Rule-based — add regex patterns for known noise (extensions, analytics, etc.)
One-off dismiss — quickly silence a specific pattern
Categories: console, network, WebSocket
Pre-built rules filter common browser extension noise

Reproduction Scripts

User actions → runnable Playwright scripts:

Multi-strategy selectors (data-testid > aria > role > CSS)
Click, input, scroll, keyboard, select, navigate events
Base URL rewriting for environment portability
Error context embedding
Configurable last-N actions scope

Test Generation

Full Playwright test scripts with configurable assertions:

Network assertions — wait for responses, validate status codes
Response shape assertions — verify JSON property structure
Console error detection — assert no unexpected errors
Custom test names and base URL substitution

PR Summaries

Performance impact reports for pull requests:

Before/after comparison table (Load Time, FCP, LCP, CLS, Bundle Size)
Delta and percentage change calculations
Error tracking (fixed vs new errors)
Compact one-liner format for git hook annotations

HAR Export

Standard HTTP Archive format export:

Filter by URL substring, HTTP method, status code range
Compatible with Chrome DevTools, Charles Proxy, and other HAR viewers
Includes request/response headers and bodies

Accessibility Audits

WCAG compliance checking with standard output:

axe-core engine with configurable rule tags
Scope to specific page sections via CSS selectors
Cached results with force-refresh option
SARIF export — standard Static Analysis Results Interchange Format
File save or inline JSON return

Persistent Session Memory

Key-value store that persists across tool calls:

Namespaced storage (save, load, list, delete)
Session context loading from previous runs
Stats and usage reporting

Extension Controls

The popup lets you dial the heat:

Setting	Options
Capture level	Errors Only · Warnings+ · All Logs
WebSocket	Lifecycle only · Include messages
Network waterfall	On / Off
Performance marks	On / Off
User actions	On / Off
Screenshot on error	On / Off
Source maps	On / Off
Domain filters	Allowlist specific sites