RVTDocs MCP

Python MCP server for deterministic RVTDocs querying with low-retry behavior.

Overview

• Python 3.14 project managed with uv
• MCP server entrypoint: rvtdocs-mcp
• Runtime source lives under src/
• Internal docs live under docs/

Tools

Tool	Purpose
rvtdocs_fetch	Primary lookup — query, year, mode (trust/full/diagnostics)
rvtdocs_scan	Lightweight existence check — no snippet
rvtdocs_debug	Full diagnostics — routing trace + extraction details
rvtdocs_batch	Concurrent multi-query (up to 10)
rvtdocs_stats	Local telemetry summary (last N hours)
rvtdocs_version_info	Supported years, default year, cache stats
rvtdocs_session_set	Store key-value in session (max 100 entries, LRU eviction)
rvtdocs_session_get	Retrieve session value

Design Goals

• Single-fetch discipline: one HTTP request per query (retries only on network failure)
• Structured extraction: selectolax HTML parsing reduces output ~33% vs raw text
• Mode-based output shaping: trust mode returns metadata only (~300 chars), full mode returns snippet (~2500 chars)
• Batch amortization: up to 10 queries per MCP call

Note: Token savings are character-based estimates. Actual LLM token counts depend on model tokenizer.

Run local

uv run --python 3.14 rvtdocs-mcp

Module form while editing source:

uv run --python 3.14 python -m src.server

Run via GitHub

uvx --from "git+https://github.com/trgiangv/rvtdocs-mcp.git" --python 3.14 rvtdocs-mcp

MCP config

From local checkout:

{
	"mcpServers": {
		"rvtdocs": {
			"type": "stdio",
			"command": "uv",
			"args": ["run", "--python", "3.14", "rvtdocs-mcp"]
		}
	}
}

From GitHub:

{
	"mcpServers": {
		"rvtdocs": {
			"type": "stdio",
			"command": "uvx",
			"args": [
				"--from",
				"git+https://github.com/trgiangv/rvtdocs-mcp.git",
				"--python",
				"3.14",
				"rvtdocs-mcp"
			]
		}
	}
}

More docs

• Setup and MCP wiring: docs/setup.md
• Internal architecture: docs/architecture.md
• Tool reference: docs/tools/
• Agent-oriented commands and repo notes: AGENTS.md