Uno Platform Separates Model Context Protocol Servers for Context and UI Verification
Uno Platform has detailed a dual Model Context Protocol (MCP) server architecture for .NET development. By separating a hosted, HTTP-based documentation server from a local stdio app execution server, agents can both inspect framework rules and verify UI rendering via Hot Reload.
Impact: Medium
Why it matters
Developers can adopt dual-tier MCP architectures to separate static rule retrieval from live execution and UI verification while managing agent context token overhead.
TL;DR
- 01Separate hosted stateless MCP documentation tools from local stateful application control tools.
- 02Monitor MCP tool definition token costs—descriptions consume context budget before any prompt text is processed.
- 03Use XML visual tree snapshots and automation peers rather than raw coordinate clicking for agentic UI debugging.
Key facts
- Docs MCP Token Cost
- ~6.4k tokens (self-reported)
- App MCP Token Cost
- ~1.5k tokens (self-reported)
- GitHub MCP Token Cost
- ~5.2k tokens (self-reported)
- Docs MCP Transport
- HTTP
- App MCP Transport
- stdio
Dual MCP Topology: Hosted HTTP vs. Local Stdio
Uno Platform splits its Model Context Protocol implementation into two targeted tiers based on lifecycle and state requirements: 1. Docs Server (`https://mcp.platform.uno/v1`): Hosted over HTTP with OAuth. It exposes tools like uno_platform_docs_search, uno_platform_docs_fetch, and /new or /init prompts to provide up-to-date framework guidance without shipping NuGet package updates. 2. App Server (Local .NET Tool): Runs over stdio directly on the developer's machine as a bridge to the Uno DevServer.
Token Overhead of Tool Definitions
Tool descriptions act as permanent prompts that consume context tokens on every request. Uno Platform reports context costs per server:
- Docs MCP Server: ~6.4k tokens
- App MCP Server: ~1.5k tokens
- GitHub MCP Server: ~5.2k tokens
Writing concise, high-signal tool descriptions is critical. For example, tool prompts explicitly direct models to prefer uno_app_element_peer_action over uno_app_pointer_click to avoid brittle coordinate clicking across different DPI settings.
Visual Tree Inspection and Self-Verification
The app server gives agents Playwright-style capabilities for native desktop/mobile apps. The agent executes uno_app_start with Hot Reload, takes pixel screenshots (uno_app_get_screenshot), parses structural layout via uno_app_visualtree_snapshot, and checks bridge connectivity using uno_health.
Try it in 2 minutes
{
"mcpServers": {
"uno_docs": {
"url": "https://mcp.platform.uno/v1"
}
}
}json
✓ When to use
- You are building or architecting custom MCP tool servers and need to optimize context token spend.
- You build cross-platform .NET applications and want agentic visual debugging with Hot Reload.
- You want to separate stateless documentation lookups from stateful local app runtime controls.
✕ When NOT to use
- You only write simple scripts without UI elements or complex tool integrations.
- You do not use Model Context Protocol (MCP) in your local IDE setup.
What to do today
- Connect your coding agent (Claude Code, Cursor, Copilot) to the hosted documentation MCP endpoint at https://mcp.platform.uno/v1.
- Review existing MCP tool prompt descriptions in your agent workflows to trim context token overhead.
- Implement structured element tree inspection tools rather than relying solely on visual screenshots for UI automation agents.
Sources