
Stylespeak
io.github.PatrizzosPublished by GitHub account Patrizzosv1.2.0Updated Oct 24, 2026
MCP server and CLI that makes CSS legible to AI agents — explains what applies and why.
Overview
Local MCP server and CLI that makes CSS legible to AI agents: it resolves the cascade, traces properties, previews change impact and explains what applies and,
- What you can do
- Resolve which CSS properties apply to a selector across given files, including the winning rule and overridden rules with confidence levels, and resolve var() chains, fallbacks and media-context values. Trace every rule that sets a property, group competing rules and show the cascade chain. Preview the blast radius of a proposed selector or property change, including shielded selectors, variable and inheritance downstream effects. Build a compressed project-wide style manifest of selectors, properties, variables, competition groups and risk hotspots. Query a running Chromium browser over CDP for exact computed styles and matched rules. Analyze plain CSS, SCSS, CSS Modules and statically-extr
- When to use it
- Useful when an AI coding agent edits styles and needs to understand the cascade before changing a rule, or when you want to check which rule wins, where a property is set, or what a change would affect. Also useful for loading a project-wide CSS summary once per session.
- Requirements
- Runs as a local stdio process; the npm package @patrizzos/stylespeak requires Node.js 21+. The live_resolve tool additionally needs a Chromium browser started with remote debugging on port 9222. No accounts, API keys or environment variables are declared.
Installation
In SourceWeft
- Open Stylespeak in the dashboard and add it to a workspace.
- Enable the server for the chats that should use its tools.
Desktop only via STDIO. STDIO servers start a local process, so they need the SourceWeft desktop host.
Other MCP clients
Follow the launch instructions in the repository.
README
StyleSpeak - MCP and CLI
A companion MCP server and CLI to stylesafe that makes CSS legible to AI agents — resolving cascade, tracing properties, predicting change impact, and explaining what applies and why before an agent touches a single line of styles.
The problem
AI coding agents write CSS without being able to see its effect. They modify a rule, assume it worked, and move on — unaware that a higher-specificity rule elsewhere already overrides it, a combinator rule in another file is silently winning, or a CSS variable resolves to something entirely different than expected. stylespeak gives agents a structured knowledge layer to consult before making changes.
What it does
resolve_styles — answers "what CSS actually applies to this selector?"
Given a selector and a set of files, returns every CSS property the selector would receive, which rule wins for each, and which rules were overridden — with confidence levels since no real DOM is available. CSS custom properties (var()) are resolved to their actual values, including chained variables, fallbacks, and media-context overrides.
trace_property — answers "everywhere this property is set, who wins?"
Given a property name and a set of files, returns every rule that sets it, groups competing rules that target overlapping selectors, and shows the full cascade chain for each group — with resolved variable values included.
impact_preview — answers "if I change this, what else breaks?"
Given a selector, property, and optional new value, predicts the full blast radius of the change before it's made — showing which selectors will see a different value, which are shielded by higher specificity, which cascade relationships are uncertain, and which downstream rules are affected through CSS variable chains or property inheritance. Works without a browser.
style_manifest — answers "what does this entire project's CSS look like at a glance?"
Builds a compressed, structured summary of a project's entire CSS knowledge — selectors, properties, variables, competition groups, and risk hotspots — that an agent can load once at the start of a session and keep in context instead of repeatedly querying individual files.
live_resolve — answers "what does the browser actually compute for this selector?"
Queries a running Chromium browser via CDP and returns exact computed styles and matched rules. No heuristics, no confidence levels — the browser resolved it. Requires Chrome running with --remote-debugging-port=9222.
Works on more than plain CSS. The four static tools above analyze .css, .scss, CSS Modules, and — as of v1.3 — styled-components and Emotion source files (.js / .jsx / .ts / .tsx), all through the same cascade engine.
Quick start
As a CLI tool
As an MCP server
Add to your MCP client config (Cursor: .cursor/mcp.json, VS Code: .vscode/mcp.json):
Once connected, agents can call:
resolve_styles({ selector, files, projectRoot, componentFiles? })trace_property({ property, files, projectRoot })impact_preview({ selector, property, newValue?, files, projectRoot })style_manifest({ files?, projectRoot?, maxSelectors? })live_resolve({ selector, port?, tabUrl? })
Example output
impact_preview
resolve_styles
Confidence levels
CSS custom property resolution
As of v0.2, stylespeak fully resolves CSS custom properties (var()) in all output:
value— the raw value as written (var(--color-primary))resolvedValue— the actual resolved value (#2563eb)variableChain— the full resolution path, including chained variablesconditionalValues— media-context overrides where the variable resolves differently
Supported: simple, fallback, nested fallback, chained, scoped, media-context, circular reference protection.
CSS-in-JS support (styled-components & Emotion)
As of v1.3, stylespeak statically extracts styles from .js, .jsx, .ts, .tsx, .mjs and .cjs files that import styled-components, @emotion/styled, @emotion/react or @emotion/css, and runs them through the same cascade engine as plain CSS. resolve_styles, trace_property, impact_preview and style_manifest all work on them, and CSS-in-JS and regular CSS can be analyzed together.
Pass JS/TS files in files, or point projectRoot at a directory — files that import one of these libraries are discovered automatically and all other JS/TS files are ignored.
Addressing components
Generated class names are hashed at runtime, so each styled component is addressed by its variable name as a class selector:
Rules are tagged cssInJs, library and component, and are treated as locally scoped (like CSS Modules): two components named Wrapper in different files are independent, never reported as conflicting, and never counted as a "declared in N locations" risk. Querying such a name is flagged as ambiguous, because results will mix both files.
What is extracted
What is flagged, never guessed
Values that depend on props or theme can't be known statically. stylespeak keeps the original expression, marks it, and tells the agent:
dynamic: trueon a declaration, winner and property means the value depends on runtime props/theme — the winner may differ depending on props.- Conditional blocks and mixin calls (
${p => p.active && css…}) can't be evaluated. They are dropped and counted indynamicInterpolations, andagentNotewarns that additional styles may apply at runtime. - Style blocks that can't be attached to a selector are listed in
cssInJs.skippedinstyle_manifestoutput instead of being silently ignored. styled(Base)recordsextends;agentNotetells the agent to also resolve the base component.
Limitations
This is static analysis, so there are boundaries:
styled(Base)does not merge the base component's styles into the child — resolve the base selector separately.- Constants and theme objects imported from other files are not resolved (same-file only).
ThemeProviderthemes are not resolved. - Emotion's
cssprop on JSX elements and style blocks not assigned to a named variable are anonymous — there is no selector to target, so they are reported as skipped. keyframesare not modeled, and.attrs({ className })is not considered.- Cross-file injection order is approximated by the order files are provided.
- For ground truth on a running app, use
live_resolveagainst the real element.
How it pairs with stylesafe
stylesafe catches problems in your CSS — conflicts, dead rules, Tailwind clashes — before they ship.
stylespeak explains your CSS — resolving cascade, tracing properties, predicting impact, resolving variables — so agents understand before they act.
Use stylesafe as a post-edit check. Use stylespeak as a pre-edit consultation. Together they give AI coding agents a complete feedback loop on styles.
Live browser inspection
live_resolve requires a Chromium browser running with remote debugging enabled:
Security note: Never run
--remote-debugging-porton a machine exposed to untrusted networks or in production. This flag opens a local API that any process on the machine can connect to.
File format support
Architecture
Zero external dependencies. Requires Node.js 21+.
Roadmap
- v0.2 ✅ — CSS custom property resolution
- v0.3 ✅ — SCSS nesting, CSS Modules scope awareness, AST component graph
- v1.0 ✅ — Chrome DevTools Protocol live resolution
- v1.1 ✅ — impact_preview: blast radius prediction before making a change
- v1.2 ✅ — style_manifest: compressed project-wide CSS knowledge for agent context
- v1.3 ✅ — CSS-in-JS: static extraction for styled-components and Emotion (template and object syntax)
Next (not yet scheduled)
Known gaps in CSS-in-JS support, in no particular order: merging styled(Base) styles into the child, resolving constants and theme objects imported from other files, and modeling keyframes and the Emotion css prop.
Source: README.md at commit 9bcac5a
Tools
0Version history
1- v1.2.0LatestSep 16, 2026
Score
50% of the applicable checks met
How complete this listing is, computed from the registry entry, the endpoint and public sources. Not a rating of the server and not a statement that it is safe or working today: reachability is a dated check, and no installs or reviews go into the score.
- Reached in the last 30 days25 pts
Keep the endpoint answering the MCP handshake; SourceWeft probes it every 30 days.
- Auth declared10 pts
Declare the credential headers in the registry entry, or keep the endpoint answering the probe so its auth can be classified.
- Tools listed15 pts
Answer tools/list without a credential, or ship a package that starts in a sandbox without one.
- DescriptionRequired15 pts
- Logo10 pts
- Documentation15 pts
- License10 pts