Context Repo

npm version Install to Cursor License: MIT

An MCP (Model Context Protocol) server that enables any MCP-compatible client to interact with your Context Repo prompts, documents, and collections — with progressive disclosure search for hierarchical document navigation.

Compatible MCP Clients

This server works with any MCP-compatible application, including:

• Claude Desktop - Anthropic's desktop app
• Cursor IDE - AI-powered code editor
• Factory Droid CLI - AI coding agent
• Any other application supporting the Model Context Protocol

What is MCP?

The Model Context Protocol is an open standard that allows AI assistants to securely connect to external data sources and tools. This server enables MCP clients to manage your Context Repo content directly.

Features

• Semantic Search - Natural language search across all your content with AI-powered relevance matching
• Prompt Management - Full CRUD with version history: list, view, create, update, delete, and restore prompts
• Document Management - Full CRUD with version history: list, view, create, update, delete, and restore documents
• Collection Management - Full CRUD: list, view, create, update, delete collections, plus add/remove items
• Version History - View and restore previous versions of prompts and documents
• Progressive Disclosure Search - Hierarchical document search with 3-level chunking (document → section → paragraph) and directional navigation
• Secure Authentication - API key-based authentication

Prerequisites

• Node.js 18 or higher
• An MCP-compatible client application
• Context Repo account with an API key

Installation

Option 1: Install globally from npm

npm install -g context-repo-mcp

Option 2: Run directly with npx

No installation needed - configure your MCP client to use npx (see configuration examples below).

Getting an API Key

1. Sign in to Context Repo
2. Go to Settings → API Keys
3. Click Create API Key
4. Select permissions:
   • prompts.read - For prompt management
   • documents.read - For document and collection access
5. Copy the key (starts with gm_)

Client Configuration

Claude Desktop

Step 1: Locate the Configuration File

macOS:

~/Library/Application Support/Claude/claude_desktop_config.json

Windows:

%APPDATA%\Claude\claude_desktop_config.json

Step 2: Open the Configuration File

macOS:

# Open in your default editor
open -e "$HOME/Library/Application Support/Claude/claude_desktop_config.json"

# Or create it if it doesn't exist
mkdir -p "$HOME/Library/Application Support/Claude"
touch "$HOME/Library/Application Support/Claude/claude_desktop_config.json"

Windows: Navigate to %APPDATA%\Claude\ in File Explorer and open or create claude_desktop_config.json.

Step 3: Add the Context Repo Server

Add the context-repo entry to your mcpServers object:

{
  "mcpServers": {
    "context-repo": {
      "command": "npx",
      "args": ["-y", "context-repo-mcp"],
      "env": {
        "CONTEXTREPO_API_KEY": "gm_your_api_key_here"
      }
    }
  }
}

If you already have other MCP servers configured, add context-repo alongside them:

{
  "mcpServers": {
    "existing-server": {
      "command": "...",
      "args": ["..."]
    },
    "context-repo": {
      "command": "npx",
      "args": ["-y", "context-repo-mcp"],
      "env": {
        "CONTEXTREPO_API_KEY": "gm_your_api_key_here"
      }
    }
  }
}

Step 4: Restart Claude Desktop

macOS: Press Cmd+Q to fully quit, then reopen Claude Desktop.

Windows: Press Alt+F4 or right-click the system tray icon and quit, then reopen.

Step 5: Verify Connection

1. Open a new conversation in Claude Desktop
2. Look for the MCP tools indicator (hammer icon 🔨) in the bottom-right of the input box
3. Click it to see "context-repo" listed with a green status

Cursor IDE

One-Click Install: Visit contextrepo.com/mcp-server and click the "Install to Cursor" button.

Manual Configuration:

Add to your Cursor MCP settings:

{
  "mcpServers": {
    "context-repo": {
      "command": "npx",
      "args": ["-y", "context-repo-mcp"],
      "env": {
        "CONTEXTREPO_API_KEY": "gm_your_api_key_here"
      }
    }
  }
}

Factory.ai Droid CLI

Factory's Droid CLI can auto-install MCP servers directly from a prompt. Simply paste this into any Droid session:

Install the following MCP Server to Droid using the NPX method.  
MCP: https://github.com/Gitmaxd/context-repo-mcp
API Key: <YOUR CONTEXT REPO API KEY>

Replace <YOUR CONTEXT REPO API KEY> with your actual API key (starts with gm_).

Droid will automatically install and configure the MCP server - no restart required.

Prefer manual configuration? You can add the Context Repo MCP server directly to your Droid MCP configuration file using the same JSON format shown in the Other MCP Clients section below.

Other MCP Clients

The general configuration pattern for any MCP client:

{
  "command": "npx",
  "args": ["-y", "context-repo-mcp"],
  "env": {
    "CONTEXTREPO_API_KEY": "gm_your_api_key_here"
  }
}

Or if installed globally:

{
  "command": "context-repo-mcp",
  "env": {
    "CONTEXTREPO_API_KEY": "gm_your_api_key_here"
  }
}

Available Tools

Once connected, your MCP client can use these tools:

Tags everywhere. As of v2.1.0, ten tools accept an optional tags: string[] argument. Filter tools (search_prompts, list_collections, list_documents, find_items) AND-match the supplied tags. Write tools (create_prompt, update_prompt, create_document, update_document, create_collection, update_collection) round-trip tags through the REST surface; on update_* tools, omit tags to leave the row untouched or send tags: [] to clear all.

User (1 tool)

Tool	Description
get_user_info	Get the authenticated user's profile details, auth method, and API-key permission scopes

Prompts (7 tools)

Tool	Description
search_prompts	List all prompts belonging to the authenticated user with optional keyword search
read_prompt	Get full prompt details including content, variables, and engine target
create_prompt	Create a new prompt template with ${variableName} syntax support
update_prompt	Update an existing prompt with automatic version history tracking
delete_prompt	Permanently delete a prompt and all its version history
get_prompt_versions	Get version history with change logs, timestamps, and version IDs
restore_prompt_version	Restore a prompt to a previous version (non-destructive, creates new version)

Documents (7 tools)

Tool	Description
list_documents	List all documents with optional collection filter and keyword search
get_document	Get full document content, title, tags, and metadata
create_document	Create a new text or markdown document with optional tags
update_document	Update a document with automatic version history and re-indexing
delete_document	Permanently delete a document, its versions, and search index entries
get_document_versions	Get version history with change logs, timestamps, and version IDs
restore_document_version	Restore a document to a previous version with re-indexing (non-destructive)

Collections (7 tools)

Tool	Description
list_collections	List all collections with names, descriptions, item counts, and IDs
get_collection	Get collection details with optional item membership list
create_collection	Create a new collection with optional color and emoji icon
update_collection	Update collection name, description, color, or icon
delete_collection	Delete a collection (items are preserved, only the folder is removed)
add_to_collection	Add documents or prompts to a collection (items can belong to multiple)
remove_from_collection	Remove items from a collection (items themselves are not deleted)

Search (1 tool)

Tool	Description
find_items	Discover prompts, documents, and collections by semantic similarity or keyword match

Progressive Disclosure (3 tools)

Tool	Description
deep_search	Search within document content returning ranked, hierarchical chunks with session deduplication
deep_read	Retrieve a single chunk with full content, hierarchy metadata, and navigation IDs
deep_expand	Navigate the document hierarchy in 5 directions (up, down, next, previous, surrounding)

Reasoning (1 tool)

Tool	Description
reason	Ask a question and get a synthesized, cited answer composed across your documents — with inline citations, an explicit gaps list, and any conflicts between sources (read-only)

Semantic Search

The find_items tool enables natural language search across your entire Context Repo. Instead of requiring exact keyword matches, it understands the meaning of your query.

How It Works

Semantic search uses AI embeddings to understand the meaning behind your query:

1. Your query is converted to a vector embedding (numerical representation of meaning)
2. Content is matched against document/prompt embeddings stored in Context Repo
3. Results are ranked by semantic similarity (relevance score 0-1)
4. Collections are scored based on the relevance of their contained items

Parameters

Parameter	Type	Required	Default	Description
query	string	Yes	-	Natural language search query
type	string	No	"all"	Filter: "prompts", "documents", "collections", or "all"
semantic	boolean	No	true	Use semantic search. Set false for literal matching
tags	string[]	No	-	Tag filter (AND-semantics: all tags must match)

Literal Mode Coverage

With semantic: false, literal matching covers titles, descriptions, and indexed document body text via the same hierarchical chunks that power deep_search (full-body matching, eventually consistent with chunking — title and preview matches are immediate). Document hits matched in the body include a highlight snippet around the exact match and a chunkId you can pass directly to deep_read or deep_expand to keep reading from that spot.

Relevance Scoring

Results are filtered by relevance score (0.0 to 1.0):

Score	Interpretation
0.7+	Excellent match - highly relevant
0.5-0.7	Good match - likely relevant
0.35-0.5	Moderate match - possibly relevant
< 0.35	Filtered out (below threshold)

Default threshold: 0.35 - Results below this score are not returned.

Example Queries

Finding related content:

"Search for my meeting notes"
"Find prompts about code review"
"What documents do I have about API design?"

Filtering by type:

"Search for 'project planning' in documents only"
"Find collection with my research materials"

Literal search (exact match):

"Search for 'README.md' with semantic disabled"

Tips for Better Results

1. Be descriptive - "prompts for writing technical documentation" works better than "docs"
2. Use natural language - Ask questions like you would to a colleague
3. Include context - "meeting notes from last sprint" is more precise than "notes"

Progressive Disclosure Search

The deep_search, deep_expand, and deep_read tools enable hierarchical document exploration. Instead of returning whole documents, progressive disclosure returns the most specific matching chunk (paragraph, section, or document level) and lets you navigate the hierarchy around it.

How It Works

Documents are organized into a 3-level hierarchy:

Document → Section → Paragraph

1. Documents are chunked into a 3-level hierarchy (document, section, paragraph)
2. Vector search finds the most specific matching level for your query
3. Navigate the hierarchy using expand directions (up, down, next, previous, surrounding)
4. Sessions track seen chunks for deduplication across iterative searches

The Three-Tool Workflow

• Step 1: deep_search — Find relevant chunks matching your query
• Step 2: deep_expand — Navigate to related content (parent sections, child paragraphs, siblings)
• Step 3: deep_read — Get full details and metadata on a specific chunk

deep_search Parameters

Parameter	Type	Required	Default	Description
query	string	Yes	-	The search query for vector similarity matching
limit	number	No	10	Maximum number of results to return
sessionId	string	No	auto-created	Session ID for result deduplication across searches
collectionId	string	No	-	Filter results to a specific collection
documentId	string	No	-	Filter results to a specific document

deep_expand Parameters

Parameter	Type	Required	Default	Description
chunkId	string	Yes	-	The chunk ID to expand from (from deep_search or deep_expand results)
direction	string	Yes	-	Navigation direction: up, down, next, previous, or surrounding
count	number	No	server default	Number of chunks to return

deep_read Parameters

Parameter	Type	Required	Default	Description
chunkId	string	Yes	-	The chunk ID to read (from deep_search or deep_expand results)

Direction Reference

Direction	Description
up	Get the parent chunk (paragraph → section → document)
down	Get child chunks (document → sections, section → paragraphs)
next	Get next sibling at the same level
previous	Get previous sibling at the same level
surrounding	Get nearby chunks for a context window

Auto-Session Deduplication

deep_search automatically creates a session on the first call. Subsequent searches within the same connection exclude previously returned chunks, enabling iterative refinement without seeing duplicate results. Providing an explicit sessionId overrides the auto-session behavior.

Difference from find_items

	find_items	deep_search
Results	Flat matches across prompts, documents, and collections	Hierarchical chunk results within documents
Best for	Finding which document contains something	Finding the exact paragraph or section and navigating around it
Navigation	None — returns top-level matches	Full hierarchy navigation via deep_expand

Example Queries

Finding content:

"Search for chunks about authentication"
"Find paragraphs mentioning API rate limits"

Navigating:

"Expand down from this section to see its paragraphs"
"Go up from this paragraph to see the full section"

Deep inspection:

"Read chunk [chunkId] for full details"

Tips for Best Results

1. Start with deep_search then use deep_expand to navigate the hierarchy
2. Use sessions for iterative refinement — auto-created by default, so repeated searches automatically skip already-seen chunks
3. Use deep_read when you need full metadata — section path, word count, heading text, and navigation IDs
4. Filter by collectionId or documentId to narrow scope before searching

Example Usage

Try these commands with your MCP client:

Prompts

"List all my prompts"
"Search for prompts about code review"
"Create a prompt called 'Bug Report' for documenting software bugs"
"Show me the details of prompt [ID]"
"Update prompt [ID] with a new description"
"Delete prompt [ID]"
"Show me the version history of prompt [ID]"
"Restore prompt [ID] to version [VERSION_ID]"

Documents

"List all my documents"
"Show me document [ID]"
"Create a document called 'Meeting Notes' with today's discussion"
"Update the title of document [ID]"
"Delete document [ID]"
"Show me the version history of document [ID]"
"Restore document [ID] to version [VERSION_ID]"

Collections

"What collections do I have?"
"Create a collection called 'Project Alpha' with a blue color"
"Show me collection [ID] with all its items"
"Add document [ID] to collection [ID]"
"Remove prompt [ID] from collection [ID]"
"Delete collection [ID]"

Search

"Search for documents about authentication"
"Find prompts related to code review"
"What do I have about project planning?"
"Search my collections for research materials"

Progressive Disclosure

"Search for chunks about API authentication"
"Expand down from chunk [ID] to see child paragraphs"
"Expand up from this paragraph to see the full section"
"Read chunk [ID] for full details and hierarchy metadata"
"Search for more results about authentication" (with auto-session dedup)

Troubleshooting

Server not connecting

1. Verify your config JSON is valid
2. Ensure you completely restarted your MCP client
3. Check that Node.js 18+ is installed: node --version

Authentication errors

• Verify your API key starts with gm_
• Check the key hasn't expired in Context Repo dashboard
• Ensure the key has required permissions

Test manually

CONTEXTREPO_API_KEY=gm_your_key_here npx context-repo-mcp

You should see the startup banner. Press Ctrl+C to exit.

Development

git clone https://github.com/Gitmaxd/context-repo-mcp.git
cd context-repo-mcp
npm install
npm test                                    # Run test suite (Vitest)
CONTEXTREPO_API_KEY=gm_your_key npm start

License

MIT - see LICENSE

Links

• Context Repo
• MCP Documentation
• GitHub Repository