Skip to main content
Glama

Prism

Deterministic web layout inspection and structural repository intelligence engine for AI coding agents via Model Context Protocol (MCP).

Quickstart (One-Line Universal Installer)

Install Prism with an interactive agent & scope selector (Global vs Local Project):

curl -fsSL https://raw.githubusercontent.com/whysooraj/prism/main/install.sh | bash

The installer prompts you:

  1. Scope Selection:

    • Global (system-wide: ~/.agents/skills/, ~/.claude.json, ~/.gemini/config/, etc.)

    • Local Project (installs inside the current directory: ./.agents/skills/, ./.cursor/mcp.json, etc.)

    • Both

  2. Agent Selection:

    • Auto-detects installed agents on your machine.

    • Lets you choose: [A] All detected agents, [E] All 70+ supported agents, or [C] Custom selection.

Non-Interactive / CI Flags

# Install globally for all detected agents non-interactively
curl -fsSL https://raw.githubusercontent.com/whysooraj/prism/main/install.sh | bash -s -- -g -y

# Install locally to current project only
curl -fsSL https://raw.githubusercontent.com/whysooraj/prism/main/install.sh | bash -s -- -l -y

Related MCP server: websight

The Problem

AI assistants fail at web design inspection and codebase analysis because they operate on low-signal approximations:

  1. Websites: Assistants receive either uncomputed raw HTML (missing exact CSS computed values) or downscaled screenshots (where padding, alignment drift, WCAG contrast violations, and touch targets cannot be mathematically verified).

  2. Repositories: Assistants either dump entire files into context (saturating tokens and inducing hallucinations) or rely on naive string grep (which lacks AST symbol understanding, caller/callee graphs, and route definitions).

Prism bridges this gap by providing deterministic, structured context:

  • Web: Exact CSS box model (padding, margin, rect), computed typography & color tokens, WCAG contrast audit, 4/8pt spacing validation, and vector wireframe SVGs.

  • Code: Ephemeral shallow repository acquisition, AST symbol parsing, caller/callee tracing, route mapping, and line-sliced retrieval.


Supported AI Agents & Ecosystem

Prism adheres to both the Model Context Protocol (MCP) and the open Agent Skills specification (agentskills.io pioneered by Vercel Labs), supporting 70+ developer agents:

Agent / Framework

Creator / Organization

Integration Type

Auto-Configured Path

Antigravity (agy)

Google DeepMind

Native MCP + Skill

~/.gemini/config/mcp_config.json~/.gemini/config/skills/prism/

OpenClaw

OpenClaw Community

MCP + Skill

~/.openclaw/mcp.json~/.openclaw/skills/prism/

Hermes Agent

Nous Research

MCP + Skill

~/.hermes/mcp.json~/.hermes/skills/prism/

Claude Code & Desktop

Anthropic

MCP + Skill

~/.claude.json~/.claude/skills/prism/

Cursor

Anysphere

MCP + Skill

~/.config/Cursor/User/mcp.json~/.cursor/skills/prism/

OpenCode

OpenCode Project

Native MCP + Universal Skill

~/.config/opencode/mcp.json

Pi Agent

Inflection AI

MCP + Universal Skill

~/.pi/mcp.json

Codex CLI

OpenAI

MCP + Universal Skill

~/.codex/mcp.json

Cline & Roo Code

Roo / Cline Community

MCP + Universal Skill

~/.config/Code/User/mcp.json~/.roo/skills/prism/

Devin for Terminal

Cognition AI

MCP + Skill

~/.config/devin/skills/prism/

Goose

Block

MCP + Skill

~/.config/goose/skills/prism/

Windsurf

Codeium

Native MCP

~/.codeium/windsurf/mcp_config.json

OpenHands

All-Hands AI

Universal Skill

~/.openhands/skills/prism/

Continue

Continue Dev

Universal Skill

~/.continue/skills/prism/

Universal Agents

Amp, Replit, Zed, Warp, Loaf, Kimi

Universal Skill

~/.agents/skills/prism/


Architecture

[ AI Assistant (agy / claude / cursor / opencode / pi / codex) ]
                               │
                 MCP (JSON-RPC via stdio / SSE)
                               │
                               ▼
                       [ PRISM ENGINE ]
            ┌──────────────────┴──────────────────┐
            ▼                                     ▼
      [ Web Pipeline ]                    [ Repo Pipeline ]
     • Playwright Chromium Engine        • Git Acquisition (--depth 1)
     • Box Model & Layout Tree           • AST Compiler (TS/JS/Py/Go/Rs)
     • Design Token Extractor            • Symbol Indexer & Call Graph
     • WCAG Contrast Auditor             • Route & Middleware Analyzer
     • Deterministic Wireframe SVG       • Surgical Line Slicer
     • SSRF Safety Boundary              • In-Memory Index Cache

Tools Reference

Tool

Parameters

Output

Description

inspect_website

url: stringviewport?: { width, height }includeWireframeSvg?: booleanincludeA11y?: booleanallowLocalhost?: boolean

JSON (WebInspectionResult)

Extracts layout tree, box model, design tokens, WCAG contrast score, and wireframe SVG

capture_screenshots

url: stringallowLocalhost?: boolean

PNG Base64 (Desktop & Mobile)

Captures dual viewports (1440x900 and 390x844)

index_repository

urlOrPath: stringbranch?: stringauthToken?: string

JSON (RepoIndexSummary)

Fetches repo, extracts AST symbols, builds file tree and routes

search_codebase

query: stringkind?: function|class|route...limit?: number

JSON (CodeSymbol[])

Searches symbols and endpoints with AST relevance scoring

get_file

filePath: stringstartLine?: numberendLine?: number

Formatted text with line numbers

Retrieves exact lines with line number prefix (42 | code)

trace_symbol

symbolName: string

JSON (SymbolTrace)

Traces definitions, importers, and connected HTTP route handlers

explain_architecture

None

JSON (ArchitectureSummary)

Summarizes tech stack, key configs, entry points, and directory map


Manual Configuration

If you prefer manual setup rather than the one-line installer:

git clone https://github.com/whysooraj/prism.git ~/.prism
cd ~/.prism
npm install
npm run build

Then add to your client's MCP configuration:

{
  "mcpServers": {
    "prism": {
      "command": "node",
      "args": ["/home/YOUR_USER/.prism/dist/index.js"]
    }
  }
}

Verification & Testing

Prism comes with a native self-check suite verifying SSRF boundaries, Playwright DOM parsing, and AST symbol extraction:

npm test

Security Model

  • SSRF Hardening: By default, inspect_website intercepts and aborts requests to private IPv4/IPv6 subnets (127.0.0.0/8, 10.0.0.0/8, 172.16.0.0/12, 192.168.0.0/16, 169.254.169.254, ::1) and local hostnames (localhost, *.local, *.internal).

  • Browser Sandboxing: Runs headless Chromium with isolated incognito contexts that terminate immediately upon evaluation completion.

  • Controlled Ingestion: Repositories are cloned shallowly with depth 1 into isolated user cache directories (~/.cache/context-bridge/repos/) and never shared across unauthorized boundaries.


License

MIT

Available Tools

7 tools
capture_screenshotsD

Inputs: url (required), allowLocalhost.

explain_architectureD

Inputs: no named parameters.

get_fileD

Inputs: filePath (required), endLine, startLine.

index_repositoryD

Inputs: urlOrPath (required), branch, authToken.

inspect_websiteD

Inputs: url (required), viewport, includeA11y, allowLocalhost, +1 more.

search_codebaseD

Inputs: query (required), kind, limit.

trace_symbolD

Inputs: symbolName (required).

Tool Schema Changelog

Recent tool additions, removals, and schema changes observed during successful MCP inspections.

Loading tool schema changelog…

    TDQS

    C2.1/5.0

    Scored across 7 tools

    Disambiguation4/5

    The tool names are mostly self-explanatory and target distinct actions: file retrieval, symbol tracing, architecture explanation, website inspection, screenshot capture, repository indexing, and code search. However, the absence of descriptions leaves some ambiguity between trace_symbol and search_codebase, and between inspect_website and capture_screenshots.

    Naming Consistency5/5

    All tool names follow a consistent snake_case verb_noun pattern: get_file, trace_symbol, explain_architecture, inspect_website, capture_screenshots, index_repository, search_codebase. There are no naming style deviations.

    Tool Count5/5

    Seven tools is well within the typical 3–15 range and each tool appears to cover a distinct capability for codebase analysis and website inspection. The set is scoped enough to avoid redundancy while covering core workflows.

    Completeness4/5

    Core read-only operations for codebase exploration and website inspection are present, including indexing, search, file retrieval, symbol tracing, and architecture explanation. Minor gaps exist, such as no explicit directory listing or index status tool, but agents can likely work around these via search.

    Maintenance

    ActivityMaintained
    ResponsivenessNo recent community issues

    Related MCP Connectors

    Related MCP Servers

    • A
      license
      Not graded
      quality
      D
      maintenance
      Enables AI agents to see, analyze, and visually verify web page changes through pixel-perfect diffing, theme extraction, layout analysis, and interactive element detection.
      MIT
    • A
      license
      A
      quality
      B
      maintenance
      Enables AI coding agents to see, measure, and verify web pages through a real Chrome browser, including screenshots, responsive layout and accessibility audits, pixel diffing against baselines, secure logins, and deployed-fix verification.
      27
      1
      MIT
    • A
      license
      Not graded
      quality
      B
      maintenance
      Enables AI agents to inspect web pages deterministically as token-cheap visual geometry instead of pixel dumps, surfacing layout tooling for overlaps, overflow, WCAG contrast ratios, and quality scoring. Agents can pass HTML source or a live URL per call and receive byte-identical, actionable findings they can act on and re-verify.
      1
      MIT