Skip to content

Fuel Any Agent

Gasoline is an open-source MCP server that implements the Model Context Protocol — a standard for connecting AI assistants to external tools. Any MCP-compatible tool can tap into your browser state. Zero dependencies. Localhost only.

ToolConfig LocationGuide
Claude Code.mcp.json (project root)Setup →
Cursor~/.cursor/mcp.jsonSetup →
Windsurf~/.codeium/windsurf/mcp_config.jsonSetup →
Claude DesktopOS-specific config fileSetup →
Zed~/.config/zed/settings.jsonSetup →
VS Code + Continue~/.continue/config.jsonBelow

Gasoline runs as a dual-mode server by default:

  • HTTP server — background process for the browser extension
  • MCP protocol — stdio channel for your AI tool
  • Auto-managed — your AI tool starts and stops the server

Gasoline exposes 5 composite tools — each with multiple sub-modes controlled by a single parameter. This reduces decision space by 79% compared to individual tool calls.

ToolWhat it doesSub-modes
observeReal-time browser stateerrors, logs, network, websocket_events, websocket_status, actions, vitals, page
analyzeData analysis and insightsperformance, api, accessibility, changes, timeline
generateCode generation from captured datareproduction, test, pr_summary, sarif, har
configureSession and noise managementstore, noise_rule, dismiss, clear
query_domLive DOM inspectionCSS selector queries

Monitor browser state in real time. Use the what parameter to select:

ModeReturns
errorsConsole errors with deduplication and noise filtering
logsAll console output (configurable limit)
networkHTTP requests — filter by URL, method, status range
websocket_eventsWebSocket messages — filter by connection ID, direction
websocket_statusActive WebSocket connections and stats
actionsUser interactions (click, input, navigate, scroll, select)
vitalsCore Web Vitals — FCP, LCP, CLS, INP
pageCurrent page URL, title, viewport

Derive insights from captured data. Use the target parameter:

TargetReturns
performancePerformance snapshot with regression detection against baselines
apiAuto-inferred API schema (OpenAPI stub or gasoline format)
accessibilityWCAG audit results with caching
changesCompressed state diff since a named/auto checkpoint
timelineMerged session timeline (actions + network + errors)

Generate code artifacts from your session. Use the format parameter:

FormatOutput
reproductionPlaywright script reproducing user actions
testPlaywright test with network/error assertions
pr_summaryMarkdown performance impact summary
sarifSARIF accessibility report (standard format)
harHTTP Archive export

Manage session state and noise filtering. Use the action parameter:

ActionEffect
storePersistent key-value storage (save/load/list/delete/stats)
noise_ruleManage noise filtering rules (add/remove/list/auto_detect)
dismissQuick one-off noise dismissal by regex pattern
clearClear all browser logs

If port 7890 is occupied:

{
"mcpServers": {
"gasoline": {
"command": "npx",
"args": ["gasoline-mcp", "--port", "7891"]
}
}
}

Update the extension’s Server URL in Options to match.

Add to ~/.continue/config.json:

{
"experimental": {
"modelContextProtocolServers": [
{
"transport": {
"type": "stdio",
"command": "npx",
"args": ["gasoline-mcp"]
}
}
]
}
}
  1. Restart your AI tool
  2. Gasoline server ignites automatically
  3. Extension popup shows “Connected”
  4. Ask your AI: “What browser errors do you see?”