| name | context-repo-mcp |
| description | Search, retrieve, version, and persist AI prompt templates, documents, and collections in the user's Context Repo account (https://contextrepo.com) via the official MCP server. Use this skill whenever the user mentions their saved prompts, prompt library, prompt templates, system prompts, "context repo", contextrepo, or wants to find, save, version, restore, or share a reusable prompt or document across their AI tools — even when they don't explicitly say "skill" or "MCP". Trigger on phrases like "find my prompt for ...", "use the system prompt I saved as ...", "save this as a reusable prompt", "what prompts do I have about ...", "version this prompt", "restore the previous version of ...", "search my context repo for ...", or "sync this prompt to my Cursor / Claude / ChatGPT setup". Also trigger when the user wants to navigate large documents hierarchically (deep_search -> deep_read -> deep_expand) or perform semantic search across their stored knowledge base.
|
| version | 2.2.2 |
| license | MIT |
| homepage | https://contextrepo.com |
| repository | https://github.com/Gitmaxd/context-repo-mcp |
| package | context-repo-mcp |
| runtime | node |
| engines | {"node":">=18.0.0"} |
| auth | ["api-key","oauth2"] |
| tags | ["mcp","prompts","prompt-management","knowledge-base","rag","claude","cursor","factory-droid","chatgpt-apps"] |
context-repo-mcp
Connect to a Context Repo account and act
on the user's saved prompts, documents, and collections through 27
MCP tools. The Context Repo platform stores reusable prompt
templates with version history, markdown documents with hierarchical
search, and collection-based organization — this skill is the
agent's read/write interface to that storage.
Capabilities
Use these tool families for the corresponding intents:
- Prompt retrieval —
search_prompts, read_prompt,
find_items (semantic + literal cross-content search).
- Prompt creation and versioning —
create_prompt,
update_prompt, get_prompt_versions, restore_prompt_version,
delete_prompt.
- Document storage —
list_documents, get_document,
create_document, update_document, delete_document,
get_document_versions, restore_document_version.
- Collection organization —
list_collections, get_collection,
create_collection, update_collection, delete_collection,
add_to_collection, remove_from_collection.
- Hierarchical document navigation —
deep_search -> deep_read
-> deep_expand (vector-backed chunk retrieval with parent /
sibling navigation across up, down, next, previous,
surrounding directions).
- Reasoning —
reason (synthesized, cited answer composed across
document content, with explicit gaps and conflicts; read-only).
- Account info —
get_user_info.
Tool inventory (27 tools, snake_case)
| Category | Tools |
|---|
| User | get_user_info |
| Prompts | search_prompts, read_prompt, create_prompt, update_prompt, delete_prompt, get_prompt_versions, restore_prompt_version |
| Documents | list_documents, get_document, create_document, update_document, delete_document, get_document_versions, restore_document_version |
| Collections | list_collections, get_collection, create_collection, update_collection, delete_collection, add_to_collection, remove_from_collection |
| Cross-search | find_items, deep_search, deep_read, deep_expand |
| Reasoning | reason |
The hosted web MCP at https://contextrepo.com/mcp additionally
exposes the OpenAI Apps SDK aliases search and fetch (29 tools
there); this npm / stdio package ships the 27 tools above.
Read-only tools are annotated with readOnlyHint: true so safe-mode
agents can execute them without write confirmation.
Output format
Every tool returns an MCP CallToolResult of shape
{ content: [{ type: "text", text: <string> }] }. The text payload
format depends on the tool family — agents must dispatch parsing
accordingly:
- Read / list / create / update tools (most tools, e.g.
read_prompt, search_prompts, list_documents, get_collection,
update_prompt, add_to_collection) return stringified JSON
(pretty-printed, 2-space indent). Parse with
JSON.parse(result.content[0].text) to consume.
- Delete tools (
delete_prompt, delete_document,
delete_collection) return a plain success or no-op string:
"✓ Deleted prompt <id>" or "<X> was already deleted (no-op).".
Do not attempt to JSON-parse these.
- Search and navigation tools (
find_items, deep_search,
deep_read, deep_expand, get_prompt_versions,
get_document_versions) return human-readable markdown
intended for direct surface to the user. IDs needed for follow-up
calls are embedded as bullet fields like **chunkId:** <id> or
**id:** <id> — extract via regex when chaining calls.
- Errors return
"Error: <message>" in the text block.
Installation
npx context-repo-mcp
Or via Claude Desktop's claude_desktop_config.json:
{
"mcpServers": {
"context-repo": {
"command": "npx",
"args": ["-y", "context-repo-mcp"],
"env": { "CONTEXT_REPO_API_KEY": "gm_..." }
}
}
}
Generate an API key at
https://contextrepo.com/dashboard/settings.
Authentication
Two equivalent auth modes:
- API Key (recommended for local agents): send
Authorization: API-Key gm_.... Generate at
https://contextrepo.com/dashboard/settings.
- OAuth 2.0 (recommended for hosted MCP clients): Clerk-issued
bearer token. Discover via
https://contextrepo.com/.well-known/oauth-protected-resource/mcp
(RFC 9728).
The MCP server's tools/list, initialize, and */list capability
methods are publicly callable, so agents can introspect this skill
before completing OAuth.
Rate limits
Per authenticated user, sliding window:
scrape family: 10 requests / 60sapi family: 100 requests / 60sreadonly family: 120 requests / 60s
Honor Retry-After headers on 429 responses.
Examples
Find and use a saved prompt
User input: "find my code review prompt and use it for this PR"
1. Call find_items(query="code review")
2. Parse JSON from content[0].text → extract prompt id from results
3. Call read_prompt(promptId=<id>)
4. Parse JSON from content[0].text → use returned `content` field as
the system message for the PR review.
Save a new reusable prompt
User input: "save this as my standard bug triage prompt"
1. Call create_prompt(
title="Bug triage",
description="Standard bug triage workflow",
content=<the prompt body>,
engine="claude-3"
)
2. Parse JSON from content[0].text → confirm new prompt id to user.
Navigate a large document hierarchically
User input: "summarize chapter 5 of my onboarding doc"
1. Call deep_search(query="chapter 5 onboarding")
→ returns markdown listing chunks; extract chunkId via
`**chunkId:** <id>` line.
2. Call deep_read(chunkId=<id>)
→ returns markdown chunk detail with hierarchy.
3. Call deep_expand(chunkId=<id>, direction="down")
→ returns child chunks for the full chapter.
4. Synthesize summary from concatenated chunk content.
Documentation
Support
