fast-html-mcp
OpenTalk2HTML-NotMD MCP Server
AI talks to you through beautiful pages — not overwhelming markdown dumps.
OpenTalk2HTML-NotMD is an MCP server that transforms how AI agents communicate with you. Instead of drowning in raw markdown files that nobody can read, AI builds clean, organized HTML pages — reports, dashboards, landing pages, invoices, anything — delivered instantly, no browser required.
The Vision
AI should communicate with humans the way humans communicate with each other: through well-designed pages, not raw text files.
Every time an AI agent dumps markdown on you, it's choosing the easiest path for the machine — not the best path for you. OpenTalk2HTML-NotMD changes that. It gives AI agents a tool to create proper, structured, beautiful HTML pages that you can actually read, scan, and share.
One command. Any AI platform. Real pages.
npx -y @aimino/fast-html-mcp-serverRelated MCP server: kiro-frontend-engineer-mcp
The Problem → The Solution
🚫 You're stuck with this | ✅ You get this instead |
Raw markdown walls you can't scan | Clean HTML pages with navigation and sections |
| Styled reports, dashboards, and data tables |
Scroll for 20 minutes to find one number | Jump between sections, read what matters |
One massive file you'll never open again | Beautiful pages you'll actually share |
AI choosing what's easy for it | AI building what's good for you |
See It In Action

AI builds a real dashboard page in seconds — no browser, no Playwright, no Docker.
Quick Start
npx -y @aimino/fast-html-mcp-serverAdd to Claude Desktop, Cursor, VS Code, or any MCP client:
{
"mcpServers": {
"open-talk-2-html-not-md": {
"command": "npx",
"args": ["-y", "@aimino/fast-html-mcp-server"]
}
}
}That's it. Your AI agent now talks to you through pages.
What Does This Look Like?
An AI agent using OpenTalk2HTML-NotMD can build you:
You ask for... | Instead of markdown | AI creates |
"Show me Q3 numbers" |
| |
"Make an invoice" |
| |
"Dashboard please" |
| |
"Pitch deck for investors" | Raw text with | |
"API documentation" | Markdown with |
See all 22+ use cases in the interactive showcase.
How It Works
OpenTalk2HTML-NotMD is a five-tier MCP server purpose-built for AI-to-human communication:
OpenTalk2HTML-NotMD MCP Server
├── Assembly → render_page (compose pages from components)
├── Patch → patch_html, set_attribute, edit_html_range (make surgical edits)
├── Read → read_html (inspect with 97% token savings)
├── Raw → write_raw_html, format_html, preview_html
├── Email → render_email, html_to_email (send to inbox)
└── Consistency→ propagate_edit, check_consistency (keep data in sync)AI builds, edits, reads, and refines HTML pages — the same way you'd work with a document, except it happens in milliseconds and costs pennies.
Why This Exists
The Markdown Trap
AI agents default to markdown because it's the easiest output format. But markdown was designed for lightweight text formatting, not for communicating complex AI outputs to humans. The result:
Users drown in walls of raw text
No structure — everything is flat, nothing is scannable
No visuals — tables, charts, and badges are ASCII at best
No sharing — you can't send a markdown file to a colleague
The HTML Alternative
HTML fixes all of this: structure, navigation, styling, visuals, sharing — built in, no browser needed on the server side. The AI creates real pages that humans can actually consume.
OpenTalk2HTML-NotMD bridges the gap: AI writes HTML instead of markdown, and you get pages you can read, scan, and share.
Tools
Tier | Tool | What It Does |
Assembly |
| Compose a page from components & template |
Assembly |
| Add your own template at runtime |
Patch |
| Replace content by CSS selector |
Patch |
| Change an attribute by CSS selector |
Patch |
| Replace lines (most token-efficient edit) |
Read |
| Inspect HTML in 4 modes (97% token savings) |
Raw |
| Write raw HTML to file |
Raw |
| Beautify HTML |
Raw |
| Preview without writing to disk |
| Build email-safe HTML | |
| Convert web HTML to email-safe | |
Consistency |
| Update all cross-references |
Consistency |
| Audit for stale references |
Utility |
| Browse available components |
Utility |
| Browse available templates |
Utility |
| See template variables |
Utility |
| See component props |
Components & Templates
22 Components
Layout: header, footer, sidebar, card-deck, grid
Data: data-table, stats-grid, timeline, financial-table, evidence-grid
Visual: risk-matrix, valuation-chart, prisma-flow
Media: figure, image-gallery
Interactive: tabs, accordion
Utility: hero, callout, code-block, citation-block
25+ Templates
Reports: report, exploration, research, code-review, equity-research, lit-review, research-briefing, scientific-paper, journal-club, earnings-summary, industry-overview
Business: invoice, budget, financial-summary, data-sheet, dashboard, financial-dashboard
Communication: newsletter, changelog, faq, meeting-notes, comparison, landing-page, error-page
Presentation: pitch-deck, deck, design, prototyping, illustrations
Developer: api-doc, custom-editor, minimal, documentation
Performance Benchmarks
Operation | OpenTalk2HTML-NotMD | Alternatives |
Cold start → first render | ~1.5s | Playwright/Puppeteer: 5-15s |
page render | ~900ms | Handlebars: similar |
Patch by #id | ~200ms | Cheerio: 2-5s |
Patch by CSS selector | ~800ms | Regex: 1-3s |
5 sequential patches | ~2s total | Re-parsing: 10s+ |
Compression ratio | 40-70% | html-minifier: 10-30% |
AI-compressed read (106KB page) | ~1,000 tokens (97% saved) | Raw: 30,553 tokens |
Why is it this fast?
#id fast-path — direct text substitution instead of full AST parsing (~10x faster)
Pre-compiled doT.js — templates compiled at startup, not render time
No browser runtime — operates on strings and AST, zero headless overhead
Architecture
The Ping-Pong Loop
1. Discover → list templates & components
2. Build → render a page
3. Inspect → read it back (97% token savings)
4. Refine → patch, edit, set attributes
5. Consistency→ propagate edits across sectionsKey Design Decisions
doT.js — 10x faster compile time than Handlebars/EJS
#id fast-path — direct string substitution for id-targeted edits
parse5 — safe AST-based HTML patching (not regex)
DOMPurify — XSS prevention on all output
Atomic writes — tmp file + rename to prevent corruption
ESM — Node.js 20+, TypeScript, ES modules
Token Efficiency (Why AI Agents Love This)
Designed from the ground up for AI token budgets:
Reading a 106KB HTML Page
Mode | Tokens | Savings | Use |
Text | 1,000 | 97% | Token-minimal reading |
Compressed | 3,909 | 87% | Summary + stats |
Content | 7,991 | 74% | Typed blocks |
Structure | 9,163 | 70% | Tree overview |
Editing (500-line page, one value change)
Approach | Tokens | When to use |
| ~2,396 | Small targets |
| ~48 | Large containers, surgical changes |
Development
git clone https://github.com/Aimino-Tech/OpenTalk2HTML-NotMD.git
cd OpenTalk2HTML-NotMD
npm install
npm run build
npm run dev # hot reload
npm test # vitestLicense
GNU General Public License v3.0 — see LICENSE.
Built by Aimino Tech — making AI communicate like a human.
Available Tools
19 toolscheck_consistencyA
Audit document for stale cross-section references and dependency violations
Inputs: mode (required), file_path (required).
edit_html_rangeA
Replace a range of lines in an HTML file with new content (most token-efficient edit — send only the changed lines)
Inputs: end_line (required), file_path (required), start_line (required), new_content (required).
edit_sectionA
Replace a rendered section in an HTML file with a new component. Pages rendered with render_page embed section markers. Use list_sections to discover available indices. Provide component+props to re-render, or html for raw replacement.
Inputs: file_path (required), section_index (required), html, props, +1 more.
format_htmlC
Format/beautify an existing HTML file
Inputs: file_path (required), options.
get_component_schemaB
Get component schema with available props
Inputs: name (required).
get_template_schemaB
Get template metadata including available variables with defaults
Inputs: name (required).
insert_sectionA
Insert a new component section at a specific index, pushing existing sections forward. Uses the same section marker system as render_page. Auto-renumbers all subsequent sections.
Inputs: component (required), file_path (required), section_index (required), html, +1 more.
list_componentsA
List available interactive components (optionally filter by category)
Inputs: category.
list_sectionsA
List all component sections in an HTML file that was built with render_page. Returns section indices, component types, and sizes.
Inputs: file_path (required).
list_templatesA
List available template shells (optionally filter by category)
Inputs: category.
patch_htmlC
Patch HTML element(s) by CSS selector
Inputs: html (required), selector (required), file_path (required).
preview_htmlB
Preview rendered HTML string without writing
Inputs: html, budget, content, options, +2 more.
propagate_editB
Propagate an entity edit through the dependency graph, updating all transitively affected sections
Inputs: file_path (required), new_value (required), old_value (required), section_selector (required).
read_htmlA
Read/analyze existing HTML file. Use mode='text' for token-efficient plain text extraction with offset/limit progressive reading.
Inputs: mode (required), file_path (required), limit, offset.
register_templateA
Register a custom template at runtime for immediate use with render_page/preview_html
Inputs: name (required), description (required), html_template (required), css, +1 more.
render_pageC
Assemble page from structured components
Inputs: sections (required), template (required), output_path (required), budget, +2 more.
set_attributeB
Set attribute on element(s) by CSS selector
Inputs: value (required), selector (required), attribute (required), file_path (required).
write_html_fileC
Write content as formatted HTML file
Inputs: output_path (required), html, budget, content, +2 more.
write_raw_htmlC
Write raw HTML content directly to file
Inputs: output_path (required), html, budget, content, +2 more.
Tool Schema Changelog
Recent tool additions, removals, and schema changes observed during successful MCP inspections.
Loading tool schema changelog…
TDQS
Scored across 19 tools
There is some overlap between tools like patch_html and set_attribute (both modify elements via CSS selectors), and write_raw_html vs write_html_file both write HTML but with unclear distinction. However, most tools have clearly defined purposes targeting different aspects (sections, templates, components), so the ambiguity is limited.
All tool names follow a consistent lower_snake_case verb_noun pattern (e.g., list_components, render_page, check_consistency). Verbs are clear and nouns align with the object acted upon, making the naming predictable and readable.
With 19 tools, the set is on the heavy side (16-25). While each tool has a specific role, the number feels slightly excessive for HTML manipulation, especially given the overlapping edit/write tools that could potentially be consolidated.
The tool surface covers creation, reading, editing, rendering, and template/component management comprehensively. The main gap is the lack of delete/removal operations (e.g., delete_element or remove_section), but core workflows are well-supported.
Maintenance
Related MCP Connectors
Deploy AI-generated HTML/CSS/JS to instant public HTTPS URLs from any MCP-compatible agent.
MCP server that lets your AI assistant build and deploy apps, hosting each one at a live URL.
AI Reasoning Cache & Consensus Layer with 11 MCP tools via Streamable HTTP.
Unified 15-tool SEO, SERP, clean Markdown & GEO (LLM citation Share-of-Voice) MCP server.
Related MCP Servers
- AlicenseNot gradedqualityDmaintenanceMCP server that enables AI coding agents to read and write to a local-first HTML/CSS design canvas, bridging visual design and code generation.1MIT
- FlicenseAqualityDmaintenanceAn MCP server implementing a 7-stage agentic frontend workflow—from design audit to PR review—including AI-driven component generation, browser validation, E2E testing, and CI self-healing.4-
- AlicenseNot gradedqualityBmaintenanceA local MCP server that lets AI agents publish markdown/HTML to a browser viewer with session management, revision diffs, and live updates via a single tool call.2MIT
- FlicenseNot gradedqualityNot gradedmaintenanceEnables AI agents to perform deterministic search synthesis and compile Sparkpage layouts into comparison matrices and ready-to-render HTML interfaces through a zero-dependency Python MCP server. Plugs into Claude Desktop, Cursor, and other MCP clients for structured, low-latency output with execution telemetry.-