cssprobe-cli MCP server

cssprobe-cli

Runtime CSS probe — inspect layout, scroll, colors, backgrounds, fonts and more in a live browser. Built for both humans and AI agents.

Static + runtime CSS diagnosis for AI agents — pair with cssgraph (the CSS knowledge graph) for the full picture.

---

Why cssprobe-cli?

When an AI agent needs to debug CSS — why is this element overflowing, what's the actual rendered layout, is this container scrollable — it can't see the browser. It guesses from static code, missing runtime-only issues like overflow, scroll chains, and containing block hijacks.

cssprobe-cli gives the agent eyes. It opens a real browser, collects computed styles and metrics via injected JavaScript, analyzes them with a confidence model, and returns structured findings — all in one command. The agent asks a question about any selector and gets back the DOM tree, ASCII layout diagram, overflow warnings, and scroll diagnosis without leaving the terminal.

---

Installation

For Humans

Copy and paste this prompt to your LLM agent (Claude Code, Cursor, Codex, etc.):

Install and configure cssprobe-cli by following the instructions here:
https://raw.githubusercontent.com/mack-peng/cssprobe-cli/main/docs/guide/installation.md

Or read the Installation Guide, but seriously, let an agent do it. Humans fat-finger configs.

For LLM Agents

Fetch the installation guide and follow it:

curl -s https://raw.githubusercontent.com/mack-peng/cssprobe-cli/main/docs/guide/installation.md

Or install directly:

npm install -g cssprobe-cli
cssprobe-cli skill-install

---

How It Works

┌───────────────────────────────────────────────────────────┐
│                      AI Agent                              │
│                                                           │
│  "Is .sidebar overflowing?"                               │
│      calls: cssprobe-cli inspect .sidebar                 │
│                             │                             │
└─────────────────────────────┬─────────────────────────────┘
                              │
                              ▼
┌───────────────────────────────────────────────────────────┐
│                  cssprobe-cli daemon                       │
│                                                           │
│  collector (IIFE) → analyzer (pure) → renderer (md/json)  │
│                             │                             │
│                             ▼                             │
│               Playwright browser (Chromium)                │
│     getBoundingClientRect · getComputedStyle · DOM walk    │
└───────────────────────────────────────────────────────────┘

1. Session — open spawns a daemon process managing a browser. Commands connect via Unix socket.
2. Collector — Injected as IIFE into the page. Walks the DOM, reads computed styles, declared values, and element metrics.
3. Analyzer — Pure Node.js functions. Classifies findings with confidence levels (DEFINITE / INDEFINITE / UNVERIFIABLE).
4. Renderer — Outputs Markdown (default) or JSON. Includes ASCII layout diagram, DOM tree, and findings.

---

Quick Start

1. Inspect a Page

# Open browser (non-blocking, starts daemon session)
cssprobe-cli open https://getbootstrap.com/docs/5.3/examples/checkout

# Inspect with CSS selector
cssprobe-cli inspect .container

# Show ASCII layout diagram
cssprobe-cli layout .container

# Show only issues/warnings
cssprobe-cli findings .container

# Close browser when done
cssprobe-cli close

2. Login-Protected Pages

# Open browser in headed mode
cssprobe-cli open https://mysite.com --headed

# User manually logs in...

# Save session state for future use
cssprobe-cli state-save --name mysite

# Reopen with saved state (no login needed)
cssprobe-cli open https://mysite.com --state state.json --headed

3. JSON Output

cssprobe-cli inspect body --json
cssprobe-cli inspect body --json | jq '.findings[] | {id, confidence, message}'

4. Use as an MCP Server (AI clients)

cssprobe-cli ships an MCP server (stdio), so AI clients call it as a native tool — no shell needed:

cssprobe-cli mcp-install          # auto-detect and configure installed clients
cssprobe-cli mcp-install --target=claude,cursor,codex
cssprobe-cli mcp-install --local  # write project-scoped config

Manual configuration for any MCP client:

{
  "mcpServers": {
    "cssprobe-cli": { "command": "cssprobe-cli", "args": ["mcp"] }
  }
}

Exposes 10 tools: cssprobe_open, cssprobe_inspect, cssprobe_tree, cssprobe_layout, cssprobe_findings, cssprobe_eval, cssprobe_screenshot, cssprobe_inject_css, cssprobe_close, cssprobe_status. Remove with cssprobe-cli mcp-uninstall.

---

Commands

Session Management

Command	Description
open [url]	Open browser in session mode (non-blocking)
open [url] --headed	Show browser window
open [url] --state <file>	Open with saved cookies + localStorage
open [url] --viewport 1280x720	Custom viewport size
open [url] --expect-url <regex>	Fail unless the final URL matches the pattern
open [url] --fail-on-login	Fail fast when redirected to a login URL
close	Close browser session
close --all	Close all sessions across all workspaces
status	Show session status (URL, title, frame count)
frames	List frames: index, name, URL, main marker
doctor	Diagnose browser install vs Playwright expected revision

CSS Inspection

Command	Description
inspect <selector>	Full CSS diagnosis (computed values, declared values, findings)
inspect <selector> --layout	Include ASCII layout diagram
inspect <selector> --brief	Compact: tree sketch + warnings/errors only
inspect <selector> --json	Structured JSON output
tree <selector>	DOM tree structure
layout <selector>	ASCII layout diagram
findings <selector>	Only issues/warnings/errors

Inspection commands accept --frame <index|name|url-substring> to target an iframe; the default is the main frame.

CSS Injection

Command	Description
inject-css <css>	Inject CSS into current page

Browser

Command	Description
resize <width> <height>	Resize browser viewport
eval <expression>	Evaluate JavaScript (browser context); DOM nodes are auto-serialized; promises are awaited, undefined fields become null
eval @file / eval -	Read the expression from a file or stdin
playwright <call>	Execute Playwright API (Node.js context)
screenshot	Take screenshot (saves to ~/.cssprobe-cli/screenshots/, --out to override)

Interaction

Command	Description
click <selector>	Real mouse click (auto-scroll + actionability wait; works with custom React components)
hover <selector>	Real mouse hover
type <selector> <text>	Fill text (replaces value); --press Enter to submit; --sequential for real key events
wait-for <selector>	Wait for selector state (--state visible|hidden|attached|detached, --timeout ms)
wait-for --js "<predicate>"	Wait until a JS predicate is truthy
wait-for --network-idle	Wait until the page is network-idle (page-level; no --frame)
wait-for ... --frame <spec>	Wait inside an iframe (selector/js modes)
sample "<expr>"	Sample an expression over time (--interval, --duration, --key, --frame); returns timeline + summary
wait-stable "<expr>"	Wait until an expression stops changing (--tolerance, --window, --timeout)
run <file>	Run a multi-step flow script

click / hover support --nth N (1-based, default first match), --force (skip actionability checks) and --timeout ms. Without --nth, multiple matches fail (strict) to prevent wrong-target clicks.

State

Command	Description
state-import <file>	Import cookies from Netscape format
state-save	Save current browser session state (cookies + localStorage)
state-check	Validate a state file: cookie count, expiry, target domain coverage (--state, --url)

Configuration

Command	Description
config-show	Show current config
config-set <key> <value>	Set config value
config-list	List all profiles
config-use <name>	Switch active profile
config-new <name>	Create new profile
config-path	Show config file path

---

Multi-step Flows

The interaction commands drive a page step by step; run turns a sequence into a versionable script:

cssprobe-cli open https://example.com/editor
cssprobe-cli wait-for .editor-canvas
cssprobe-cli click ".layout-switch" --nth 1
cssprobe-cli type ".block-title input" "Hello" --press Enter
cssprobe-cli inspect .block-grid --json
cssprobe-cli close

For repeatable flows, write a script and run it with one command. The script runs inside the daemon session (full Playwright page access):

// flow.js
export default async ({ page, context, frames, newPage, click, waitFor, type, inspect, screenshot, expect, log }) => {
  await page.goto('https://example.com/editor', { waitUntil: 'domcontentloaded' });
  await waitFor('.editor-canvas', { state: 'visible', timeout: 10000 });

  await click('.layout-switch', { nth: 1 });
  await waitFor({ js: "() => document.body.dataset.layout === 'two-col'" });

  await type('.block-title input', 'Hello', { press: 'Enter', sequential: true, clear: true });
  await waitFor({ networkIdle: true });

  const { findings } = await inspect('.block-grid', { format: 'json' });
  log('findings:', findings.length);
  await screenshot({ path: './after-layout-switch.png' });

  return { findings: findings.length };
};

cssprobe-cli run ./flow.js        # --timeout ms for the whole script (default 300000, 0 = unlimited)

Scripts must default-export an async function receiving the API object: page, context, frames(), newPage(), click, hover, type, waitFor, inspect, screenshot, expect, log. waitFor accepts { state, timeout, frame }. Uncaught page errors are collected into the result envelope as pageErrors. screenshot relative paths resolve against the script's directory; logs (log() and console.*) are collected and returned with the result. On failure the command exits 1, prints the collected logs, and keeps the session alive.

Scripts are executed by the local daemon with full Node privileges — only run trusted scripts.

---

Scenario Suites (Regression)

Run repeatable regression cases from a JSON suite: each case opens a fresh browser, executes steps, asserts on saved values, and writes artifacts + an optional JUnit report.

cssprobe-cli scenario run suites/ticket-26571394.json \
  --state ~/.cssprobe-cli/states/preprod.json \
  --artifacts out/ticket-26571394 \
  --junit out/junit.xml

{
  "name": "ticket 26571394 preprod",
  "defaults": { "viewport": { "width": 1280, "height": 800 } },
  "cases": [
    {
      "name": "refresh keeps the hash section aligned",
      "url": "https://example.com/#listen-to-tcmi",
      "steps": [
        { "do": "sample", "expression": "({ hash: location.hash, top: Math.round(document.querySelector('#listen-to-tcmi')?.getBoundingClientRect().top ?? -1) })", "interval": 250, "duration": 12000, "key": "top", "tailFrom": 9500, "save": "timeline" },
        { "do": "assert", "label": "hash retained", "on": "timeline", "path": "summary.last.value.hash", "eq": "#listen-to-tcmi" },
        { "do": "assert", "label": "anchor stable", "on": "timeline", "path": "summary.tail.min", "min": 70 },
        { "do": "assert", "label": "anchor stable", "on": "timeline", "path": "summary.tail.max", "max": 90 }
      ]
    },
    {
      "name": "deployment fingerprint",
      "url": "https://example.com/",
      "steps": [
        { "do": "waitFor", "js": "() => !!window.DEBUG?.NavigatorStore", "timeout": 20000 },
        { "do": "assert", "globals": { "DEBUG.NavigatorStore.setHashRestoreProtected": "function" }, "absent": ["DEBUG.NavigatorStore.navigateToSectionId"] }
      ]
    }
  ]
}

Steps: goto, reload, wait, eval, sample, waitStable, click, waitFor, screenshot, assert. wait takes a fixed ms; waitFor supports frame for in-iframe selector/js waits. sample and eval save values with save: "<name>"; assertions read them with on: "<name>" + path, or check live state (url, title, frames), url, globals, absent. sample supports key, tailFrom, ignoreErrors; every run writes summary.json, per-case case.json, sample series and failure screenshots.

Iframe and preview pages. A page can contain hidden helper iframes (about:blank); use frames to pick the real one, and state: "attached" when waiting for a frame element that is not visible. Preview frames can take 6–8s to render and analytics requests keep networkIdle busy — prefer a fixed wait step followed by a waitFor inside the frame (frame: "i/preview") instead of networkIdle.

Confidence Model

Every finding carries a confidence level:

Level	Meaning
DEFINITE	Based on computed values (facts from getComputedStyle) or accessible declared values
INDEFINITE	Declared value uses % — resolves at runtime
UNVERIFIABLE	Declared value missing or from blocked cross-origin stylesheet

The report header shows: confidence: DEFINITE 8 | INDEFINITE 0 | UNVERIFIABLE 1

---

Complementary Tool: cssgraph

cssprobe-cli is a runtime CSS probe — it opens a real browser and reads what is actually rendered. For static questions about the source code — where is .btn-primary defined, what cascades over it, which components reference it, what would change if I edited it — use cssgraph, a local SQLite knowledge graph of every className, property, variable, and at-rule in your stylesheets. The two tools complement each other:

	cssgraph (static)	cssprobe-cli (runtime)
Analyzes	source code (CSS/SCSS/Less + JSX/TSX + templates)	live browser (computed styles + DOM)
Answers	where a class is defined, cascade, impact, unused CSS	actual rendered layout, overflow, scroll, height chains
When	before/without a browser	after static analysis, to verify at runtime
Usage	cssgraph init + explore/rule/impact	open + inspect/layout/findings

Recommended workflow for a style problem — go static first, then verify at runtime:

# 1. Static: find the definitions, cascade, and blast radius (cssgraph)
cssgraph explore .btn-primary
cssgraph cascade .btn-primary

# 2. Runtime: confirm what actually renders (cssprobe-cli)
cssprobe-cli open https://example.com
cssprobe-cli inspect .btn-primary
cssprobe-cli findings .btn-primary

Install: npm install -g cssgraph (requires Node.js >= 22.5.0). It exposes 12 MCP tools (cssgraph_explore, cssgraph_rule, cssgraph_impact, …) plus CLI commands. See npm / GitHub.

---

Configuration

# Set values
cssprobe-cli config-set browser chromium
cssprobe-cli config-set depth 8
cssprobe-cli config-set viewport 375x812

# Profile management
cssprobe-cli config-new staging
cssprobe-cli -p staging config-set browser firefox
cssprobe-cli config-use staging

Priority: CLI flags > Environment variables > Config file

--browser       CSSPROBE_BROWSER
--depth         CSSPROBE_DEPTH
--headed        CSSPROBE_HEADED
--viewport      CSSPROBE_VIEWPORT
-p, --profile   CSSPROBE_PROFILE

Config file: ~/.cssprobe-clirc

---

Supported Platforms

Platform	Architectures
macOS	x64, arm64
Linux	x64, arm64
Windows	x64, arm64

---

Development

npm install
npm run build       # tsc + esbuild collector + daemon entry + generate help.json
npm test            # Run unit tests (58 tests)
npx tsc --noEmit    # Type check only

If cssprobe-cli is useful

A ⭐ star helps other developers discover it — and tells us to keep shipping.

---

License

MIT