System Architecture

Proxima v5.0.0 uses a modular, multi-tier agentic architecture that translates Model Context Protocol (MCP) tool requests, coordinates local/offline endpoints, and manages user browser sessions locally and privately.

1. System Topology & Data Flow

The platform consists of three decoupled operational tiers: the Electron status hub, the Node.js MCP server middleware, and the autonomous Python agent runtime.

2. Electron Hub Layer (Main Process)

The Electron shell acts as the coordinator and system host. It manages sessions, stores credentials, and starts the background browser views.

Stealth Session Compliance

For Session Routing, Proxima loads each provider inside separate BrowserView nodes. It intercepts response streams directly at the renderer engine level, bypasses traditional HTML selector scraping, and applies compliance parameters (stealth User-Agents, disabling Blink automation indicators, custom preloads) to ensure session reliability.

Key Vault Encryption

When in BYOK (Bring Your Own Key) Mode, Proxima securely encrypts and stores your keys inside your operating system's native keychain vault (via Electron's SafeStorage API). No keys are ever stored in plain text or uploaded to external servers.

3. Node.js MCP Middleware Pipeline

The MCP server process acts as the local middleware intelligence. Every query is parsed, categorized, and optimized before being sent to the LLM:

EMA Smart Router

Classifies incoming queries into categories (e.g., code generation, research, security scanning) and routes them based on provider ratings. It computes provider performance using an Exponential Moving Average (EMA) latency formula:

If a provider experiences three consecutive timeouts, the smart router automatically flags it and puts it into cooldown for 60 seconds, falling back to other providers.

Context7 Doc Middleware

Intercepts questions to detect library mentions (like React, FastAPI, or Docker). It automatically retrieves up-to-date documentation schemas from Context7's database and injects them as pre-context to prevent model hallucinations.

4. Autonomous Agent Integration Layer

Proxima includes a decoupled Python runtime daemon that connects tool calls directly to physical system operations. This daemon listens on local port 8500 (guarded against cross-origin scripting) and executes self-healing run loops, CDP browser interactions, and workflow tasks.