# ServiceNow Docs MCP Server

An MCP server that gives AI agents fast, reliable access to the full ServiceNow documentation catalog.

## Role

A retrieval layer purpose-built for AI agents working on the ServiceNow platform.

## Stack

- Python
- FastMCP
- ChromaDB
- HuggingFace
- Fly.io

## Metrics

- **287,271 — chunks indexed:** Full ServiceNow docs ingested into ChromaDB.
- **231 — bundle filters:** Scope searches to a specific product area or release.
- **4 — MCP tools:** Search, list bundles, reassemble full docs, and health stats.

## Architecture

1. **Ingest:** Downloads the ServiceNow documentation dataset, generates embeddings, and loads the full corpus into ChromaDB.
2. **Index:** Each chunk is stored with its product area, title, URL, and position so agents can filter results or retrieve the full source page.
3. **Serve:** FastMCP exposes four tools: search, product-area discovery, full-page retrieval, and health.
4. **Operate:** Runs on Fly.io with a persistent volume, bearer-token auth, and a startup warmup to absorb cold-start latency.

## Problem

AI agents working on ServiceNow tasks need documentation that is both broad and trustworthy. Live scraping is fragile and slow; general web search is unreliable.

ServiceNow currently does not have an official MCP Server that serves up product documentation.


## Features

- Tools: semantic search, bundle discovery, full-document reconstruction, health reporting.
- Runs locally with zero config or accessible from any MCP-compatible client.
- Tests cover tool behavior, auth middleware, startup validation, and ingest logic.

## Retrieval Quality

A simple evaluation benchmarks the index against representative ServiceNow tasks such as incident management, Flow Designer error handling, GlideRecord usage, and Event Management rules.

Average top-1 score of 0.741


## Links

- [Demo](https://sn-docs-mcp-demo.vercel.app/)