Turn any documentation into an AI-searchable knowledge base. Feed it markdown, HTML, PDFs, web pages, or plain text — publish to a database and let anyone on your team query it through AI tools like Claude, Cursor, or Slack. Powered by the Model Context Protocol.
What This Does
AI-powered company knowledge MCP. Unified place for internal policies, values, documentation, and governance. Agents can search, cite, and answer questions using real company docs.
星标
49
工具
—
可用率
—
Ingest — Point the CLI at your content. It accepts markdown files, HTML pages, PDFs, live URLs, or even an entire website via crawl.
Publish — Run a command that converts your content into searchable vectors and stores it in a database.
Query — Connect any MCP-compatible AI tool to your server. Ask questions in plain English and get answers sourced directly from your documentation.
Two Ways to Use This
There are two distinct roles when working with Company Docs MCP. Most people on your team only need the first one.
If someone already set up the server for your team
You just need a URL. No accounts, no installation, no terminal commands.
Get the server URL from whoever set it up (it looks like https://company-docs-mcp.example.workers.dev/mcp)
Cursor / Windsurf: Add the URL as a remote MCP server in settings
Start asking questions about your documentation
That's it. Cloudflare, Supabase, and the CLI are only needed by the person who sets up and maintains the server.
If you're setting up the server (admin/maintainer)
The rest of this README is for you. Follow the setup guide below to get everything running.
How the Pieces Fit Together
The system uses three services. All three offer free tiers that are sufficient for most teams.
flowchart TD
A["Your Content<br/>Markdown, HTML, PDFs, URLs"]
B["Cloudflare Workers AI"]
C[("Supabase")]
D["Your Team<br/>Claude, Cursor, Slack, Chat UI"]
E["Cloudflare Worker"]
A -- "ingest + publish" --> B
B -- "store vectors" --> C
D -- "ask a question" --> E
E -- "vector search" --> C
C -. "matching docs" .-> E
E -. "answers" .-> D
style A fill:#f9f9f9,stroke:#333,color:#333
style B fill:#dbeafe,stroke:#1d4ed8,color:#333
style C fill:#d4edda,stroke:#155724,color:#333
style D fill:#f0fdf4,stroke:#15803d,color:#333
style E fill:#dbeafe,stroke:#1d4ed8,color:#333
Service
What it does
Why it's needed
Cloudflare
Hosts your server and converts text into searchable vectors using its built-in AI
This is where your server runs 24/7 so your team can query docs at any time. It also handles the AI processing that makes semantic search possible — no separate AI subscription needed.
Supabase
Stores your documentation in a PostgreSQL database with vector search
Powers "smart" search — asking "how do I deploy?" will find documents about releases, CI/CD, and shipping, not just pages containing the word "deploy."
npm package
A command-line tool that ingests your content (markdown, HTML, PDFs, URLs) and publishes it to the database
You run this on your computer whenever you add or update documentation.
No third-party AI API keys are required. Cloudflare provides the AI capabilities through its Workers AI service, which is included with every Cloudflare account at no extra cost.
What You'll Need
Before starting, create free accounts on these two services:
Node.js 18 or later — the runtime that powers the CLI tool (download here)
A Cloudflare account — for hosting and AI (sign up here, free tier works)
A Supabase account — for the database (sign up here, free tier works)
That's it. No OpenAI, Anthropic, or Google API keys needed.
Setup Guide
Follow these steps in order. Each one builds on the previous.
Step 1: Install the Package
Open your terminal in the project where your documentation lives and run:
npm install company-docs-mcp
This downloads the CLI tool to your project. No external services are contacted yet.
Step 2: Create Your Database (Supabase)
Your documentation needs a database to store content and make it searchable.
Go to Settings > API and copy three values (you'll need these in Step 4):
Project URL (looks like https://abc123.supabase.co)
anon key (a long string starting with eyJ)
service_role key (another long string starting with eyJ — keep this private)
Open the SQL Editor in the left sidebar, paste the contents of database/schema.sql, and click Run
This creates the database tables and search functions the system uses.
The schema file is included in the npm package at node_modules/company-docs-mcp/database/schema.sql.
Step 3: Log In to Cloudflare
The CLI needs access to Cloudflare's AI service to convert your documentation into searchable vectors. The simplest way to connect is through the Wrangler CLI (Cloudflare's command-line tool, included with this package).
Run this command:
npx wrangler login
A browser window will open asking you to log in to your Cloudflare account and grant permission. Click Allow and return to your terminal.
Your Account ID is shown on the right side of the overview page — copy it
That's the only Cloudflare setup needed for publishing. The CLI automatically detects the login credentials that wrangler login saved to your computer.
Token expiration: The login session expires periodically. If you see an authentication error when publishing, just run npx wrangler login again.
Step 4: Configure Your Environment
Create a file called .env in your project root with these values:
# Supabase — where your documentation is stored
SUPABASE_URL=https://your-project.supabase.co
SUPABASE_ANON_KEY=eyJ...
SUPABASE_SERVICE_KEY=eyJ...
# Cloudflare — your Account ID (from Step 3)
CLOUDFLARE_ACCOUNT_ID=your-account-id
Replace the placeholder values with the ones you copied from Supabase (Step 2) and Cloudflare (Step 3).
Keep this file private. Never commit .env to version control — it contains credentials. Add .env to your .gitignore file.
Step 5: Prepare Your Content
The system accepts multiple content formats. Use whichever fits your workflow:
Format
How to ingest
Best for
Markdown (.md)
npx company-docs ingest markdown --dir=./docs
Documentation you write and maintain as files
HTML (.html)
npm run ingest -- html ./file.html
Exported web pages, saved articles
PDF (.pdf)
npm run ingest -- pdf ./document.pdf
Policies, reports, specs in PDF format
URL (any web page)
npm run ingest -- url https://example.com/page
Live web content you want to index
Website crawl
npm run ingest:web -- --url=https://docs.example.com
Entire documentation sites
CSV of URLs
npm run ingest:csv -- urls.csv
Batch-importing many web pages at once
Note: The npx company-docs CLI (from the npm package) directly supports markdown ingestion and publishing. The other formats (HTML, PDF, URL, website crawl, CSV) are available when running from the cloned repository.
Markdown is the most common starting point. Create files in a directory — any folder structure works:
You can optionally add YAML frontmatter to control how each document is categorized:
---
title: Deployment Guide
category: engineering
tags: [deploy, ci-cd, release]
description: How to deploy to production
---
# Deployment Guide
Your content here...
If you don't include frontmatter, the system will auto-detect a category and extract tags from the content.
Step 6: Ingest and Publish
Two steps turn your content into a searchable knowledge base:
Step 1 — Ingest (converts your content into structured entries):
# Markdown (most common)
npx company-docs ingest markdown --dir=./docs
# Or from the cloned repo, for other formats:
npm run ingest -- html ./exported-page.html
npm run ingest -- pdf ./policy-document.pdf
npm run ingest -- url https://wiki.example.com/important-page
Step 2 — Publish (pushes entries to the database with search vectors):
npx company-docs publish
What happens:
The ingest step reads your content, extracts titles and sections, and saves structured entries as JSON files in a content/entries/ folder in your project.
publish sends each entry to Cloudflare's AI to generate search vectors, then stores everything in your Supabase database. A content hash automatically skips entries that haven't changed, so re-running is fast.
To preview what would be published without actually writing to the database:
npx company-docs publish --dry-run
Updating content: Whenever you change your source files, run both steps again. Only changed entries are re-processed.
Step 7: Deploy the Server (Cloudflare Worker)
The server is what runs 24/7 and handles search queries from your team's AI tools. It's deployed as a Cloudflare Worker.
Clone the repository
git clone https://github.com/southleft/company-docs-mcp.git
cd company-docs-mcp
npm install
Cursor / Windsurf / Other MCP clients: Add the URL as a remote MCP server in your client's settings.
Once connected, your AI tool will have access to these search tools:
Tool
What it does
search_documentation
Finds documentation that matches your question using semantic search
search_chunks
Searches specific sections within documents
browse_by_category
Lists all documentation in a category (categories come from frontmatter, the --category flag, or auto-detection)
get_all_tags
Lists every tag used across your documentation
Cloudflare's Role — A Quick Summary
Since Cloudflare appears in several steps, here's a plain-language summary of what it does and when:
When
What Cloudflare does
How it's accessed
Publishing docs (Step 6)
Converts your text into numerical vectors that enable semantic search
CLI calls the Cloudflare REST API using your wrangler login credentials
Running the server (Step 7+)
Hosts the always-on server that your team queries; generates vectors for incoming questions
Built-in — no API keys needed at runtime
Is Cloudflare optional? No — it's required for both publishing and hosting. However, the free tier is more than sufficient and no separate AI subscription is needed. The only setup required is creating an account and running npx wrangler login.
Advanced: Pluggable Backends
The default stack above (Cloudflare Workers + Workers AI + Supabase) is the
recommended path and works out of the box. For organizations with existing
infrastructure preferences, the backends are pluggable — each one resolves
through a provider registry selected by environment variable:
Seam
Env var
Ships with
Embeddings
EMBEDDING_PROVIDER
workers-ai (default), openai
Vector store
VECTOR_STORE
supabase (default)
Chat (optional demo/Slack only)
CHAT_PROVIDER
openai (default), anthropic
Cache
CACHE
kv (default with a KV binding), memory, none
Adding a backend (e.g. a different vector database) means implementing one
interface and registering it under an id — no core code changes. See
src/providers/README.md for the contracts and
docs/EMBEDDING_MIGRATION.md before switching
embedding providers (that one requires re-embedding your content).
Note for MCP-first deployments: the chat provider only affects the optional
web chat demo and Slack integration. MCP clients bring their own language
model, so search quality depends only on the embedding provider and vector
store.
CLI Reference
npm Package Commands
These work anywhere via npx company-docs:
company-docs <command> [options]
Command
Description
ingest markdown
Parse markdown files into content/entries/
publish
Push entries to the database with AI-generated vectors
ingest supabase
Same as publish
manifest
Generate content/manifest.json (used during Worker deployment)
Category label for the content (overrides frontmatter)
documentation
--recursive
Include files in subfolders
true
--verbose, -v
Show detailed output
false
Publish Options
Option
Description
--clear
Delete all existing data before publishing (start fresh)
--dry-run
Preview what would change without writing to the database
--verbose
Show detailed per-entry progress
Examples
# Ingest markdown from different folders with different categories
npx company-docs ingest markdown --dir=./docs/engineering --category=engineering
npx company-docs ingest markdown --dir=./docs/policies --category=hr
npx company-docs publish
# Ingest a PDF and a web page (from cloned repo)
npm run ingest -- pdf ./policies/employee-handbook.pdf
npm run ingest -- url https://wiki.example.com/onboarding
npx company-docs publish
# Crawl an entire documentation site (from cloned repo)
npm run ingest:web -- --url=https://docs.example.com --max-pages=50
# Full re-publish from scratch
npx company-docs publish --clear
# Preview changes
npx company-docs publish --dry-run --verbose
YAML Frontmatter Reference
Each markdown file can optionally include a YAML frontmatter block at the very top. The system reads these fields:
---
title: Page Title
category: engineering
tags: [deploy, ci-cd, release]
description: A short summary of this page
status: stable
version: 1.0.0
source: src/path/to/source.ts
figma: https://figma.com/...
author: Jane Smith
department: Engineering
---
Field
Effect
title
Used as the document title (overrides the first # Heading)
category
Sets the browseable category for this document
tags
Adds tags for filtering and discovery
description
Stored as metadata, returned in search results
status
Stored as metadata (e.g., draft, stable, deprecated)
version
Stored as metadata
source, figma, author, department
Stored as metadata, available in search results
All fields are optional. If no frontmatter is present, the system auto-detects a category and extracts tags from the content.
Priority order: Frontmatter values take highest priority, followed by CLI flags (like --category), followed by auto-detection.
Incremental Updates
The system is designed for repeated runs — you don't need to start from scratch each time:
Content hashing — Only entries whose content has actually changed are re-processed
Deterministic IDs — The same source file or URL always produces the same database ID, preventing duplicates
Stale cleanup — Entries removed from your source files are automatically cleaned up from the database
# Edit your content, then re-ingest and re-publish — only changes are processed
npx company-docs ingest markdown --dir=./docs
npx company-docs publish
Optional: Slack Integration
The server includes a Slack slash command so team members can search documentation directly from Slack:
/docs deployment process
/docs PTO policy
/docs how to set up staging
The server includes a web-based chat UI at its root URL (visit the Worker URL in a browser). It has two modes:
Search mode — Finds relevant documentation using the same vector search as MCP. No additional setup needed.
AI chat mode — Sends your question to OpenAI GPT-4o, which searches your docs and synthesizes a conversational answer. This is the only feature that requires an OpenAI API key (set as a Worker secret: OPENAI_API_KEY).
Customize the chat UI with environment variables in wrangler.toml:
By default, the system uses Cloudflare's Workers AI for embeddings (free, no extra keys). If your organization prefers OpenAI, you can switch:
OPENAI_API_KEY=sk-...
EMBEDDING_PROVIDER=openai
Provider
Model
Dimensions
When to use
Workers AI (default)
@cf/baai/bge-large-en-v1.5
1024
Default. No extra keys. Free on Cloudflare.
OpenAI
text-embedding-3-small
1536
If your organization already standardizes on OpenAI.
Important: The embedding provider must match the database schema. The default schema.sql uses 1024 dimensions (Workers AI). If switching to OpenAI, change all vector(1024) to vector(1536) in the schema before running it.
Troubleshooting
No results from search
Verify npx company-docs publish completed without errors
Check that your .env has the correct Supabase credentials
Run npx company-docs publish --dry-run to see what entries exist
Authentication error when publishing
Your wrangler login session may have expired — run npx wrangler login again
Verify CLOUDFLARE_ACCOUNT_ID is set in your .env
Duplicate entries
Re-run npx company-docs ingest markdown followed by npx company-docs publish — duplicates are cleaned up automatically
MCP client not connecting
Make sure the Worker is deployed and accessible
Use the /mcp path in the URL (not just the root URL)
Restart your MCP client after adding the connector
Wrangler login not working
If you have a CLOUDFLARE_API_TOKEN set in your environment or .env file, it can interfere with the login flow. Remove or comment it out, then try npx wrangler login again.
Security
Never commit .env files — they contain credentials
The SUPABASE_SERVICE_KEY has full database access — keep it private
The SUPABASE_ANON_KEY is restricted by Row Level Security policies (read-only)
AI-powered company knowledge MCP. Unified place for internal policies, values, documentation, and governance. Agents can search, cite, and answer questions using real company docs.
上次检查时间:2026-10-27。 AI-powered company knowledge MCP. Unified place for internal policies, values, documentation, and governance. Agents can search, cite, and answer questions using real company docs.