Skip to content
Figranium logo
Open menu
HomeDocsReleases
Introducing Figranium MCP: Giving AI Agents Full-Cycle Browser Autonomy

Introducing Figranium MCP: Giving AI Agents Full-Cycle Browser Autonomy

Today, I am excited to announce the release of Figranium v0.13.2 alongside the official launch of the Figranium Model Context Protocol (MCP) Server.

Until now, using AI for web automation meant choosing between two extremes: writing brittle, custom Playwright scripts by hand, or relying on expensive visual models that burn through tokens just to find a single button click.

By exposing Figranium’s underlying engine—including the live Inspector, Stealth Engine, and Task Runner—directly to MCP-compliant environments like Cursor, Claude, and ChatGPT, I am bridging that gap. AI agents can now inspect, build, test, and maintain production-grade automations directly like a human engineer.

What’s New in Figranium v0.13.2#

This release brings together foundational architectural shifts, UI polish, and full programmatic accessibility:

1. Programmatic API Bridge (Exposed Endpoints)#

Figranium’s core operational tools are now fully exposed via public API routes:

  • POST /api/browser/open: Open and control browser instances dynamically.
  • GET /api/inspector/highlight: Programmatically highlight and return high-confidence DOM CSS/XPath targets from live pages.
  • PATCH & DELETE /api/tasks/:id: Modify task steps on the fly or clean up temporary workflows.
  • POST /api/tasks/run-ad-hoc: Execute ad-hoc task definitions instantly without saving them permanently.

2. UI Refinements & Typography Alignment#

I've updated the monospace typography across all task actions, JSON payloads, and dynamic logs from JetBrains Mono to Space Mono. This change complements Figranium's primary typeface, Questrial, establishing a cohesive aesthetic throughout the app.

3. Templates Hub Reset (Apify Store Model)#

Similar to the Apify Store, the Templates Hub serves as the central directory for reusable web automation tasks. I completely cleared out the existing templates list for v0.13.2. All previous tasks were created prior to v0.9—meaning they were significantly slower and failed to leverage the engine's current performance potential. Wiping them provides a clean baseline to build and feature modern, high-speed workflows.

4. Official NPM Deprecation#

Following the freeze on NPM package updates in v0.13.0, the core Figranium NPM installation model is now officially deprecated. Moving forward, Figranium should be deployed using standalone binaries, Docker containers, or direct source builds to ensure complete runtime parity.

(Note: figranium-mcp and n8n-nodes-figranium remain fully active on NPM.).

Why Figranium + MCP?#

The Model Context Protocol (MCP) gives LLMs a standardized way to invoke external tools. By plugging Figranium into MCP, your AI assistant gets native, full-cycle control over web tasks:

  • Visual & Structural Inspection: The agent uses /api/inspector/highlight to programmatically extract high-confidence CSS and XPath targets straight from live DOM states.
  • Deterministic Execution Engine: Rather than re-evaluating every step via computer vision, the agent compiles natural language intent into structured, stateful Figranium workflows complete with conditional logic (if/else), JS extractions, and REST API calls.
  • Human-like Stealth Integration: Powered by Figranium's native interaction engine, agent workflows automatically simulate human cursor glide, dynamic hovering, natural typing cadence, and subtle idle movements.
  • Autonomous Feedback Loops: Agents can execute tasks ad-hoc (task_execute), inspect returned payloads, handle dynamic errors, and delete throwaway test tasks (task_delete) without human intervention.

How It Works in Practice#

Here is an example of a single prompt given to Cursor using the Figranium MCP server:

"Inspect Hacker News, pull the top 5 stories, click into page 2, fetch additional item data via the Firebase API, and take a screenshot."

Behind the scenes, the agent executes a full Inspect-Build-Test-Clean up cycle:

  1. Inspects: Queries page elements and verifies DOM selectors.
  2. Builds: Generates a stateful workflow JSON with embedded JavaScript parsers and out-of-band HTTP calls.
  3. Executes: Runs the engine headlessly with stealth behaviors enabled (idleMovements, cursorGlide).
  4. Verifies: Reads the structured extraction payload and screenshot artifact to ensure 100% execution accuracy.

Getting Started#

You can bridge your AI environment to your local Figranium instance today using figranium-mcp:

{
  "mcpServers": {
    "figranium": {
      "command": "npx",
      "args": ["-y", "figranium-mcp"]
    }
  }
}

Add the configuration above to your Cursor or Claude Desktop settings file to instantly grant your AI agent full browser automation capabilities.

What’s Next?#

Work is underway on hosted cloud endpoints (not hosted instances—rather free endpoints just to make MCP setup easier) and one-click installers for native plugin directories across Claude Desktop, ChatGPT, and Manus.

Previous
Next