New: Ask reason for a cited answer
Context Repo

MCP server

Your context, everywhere.

Give Claude, Cursor, ChatGPT, and other MCP clients direct access to your prompts, documents, and collections. One connection, 29 tools.

Read the docs
  • 29 tools
  • Streamable HTTP
  • stdio bridge v3

~/.cursor/mcp.json

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

Works with the MCP clients you already use

  • Cursor
  • Claude
  • ChatGPT
  • VS Code
  • Windsurf
  • Amp
  • Factory
  • Grok

Say it in plain language.

Your client picks the right tools behind the scenes. You never leave the conversation.

You say

“Use my PR review prompt on this diff.”

Your client calls

  1. read_prompt

Your client fetches the review template with all of its variables and applies it to the diff in front of you. Nothing to copy.

A real Cursor agent chat in dark mode with both sidebars hidden: the request 'Use my PR review prompt on this diff.' with the branch diff attached, the collapsed 'Worked for 28s' fold, the line 'Reviewed burst-allowance vs main with your PR Review prompt.', and the start of the review in the prompt's own structure: a Summary and a Must fix list quoting file and line.

29 tools. One URL.

Full CRUD, search, Deep Search, and a cited reason answer, over authenticated Streamable HTTP or the stdio bridge.

  • User1

    • get_user_info

  • Prompts7

    • search_prompts
    • read_prompt
    • create_prompt
    • update_prompt
    • delete_prompt
    • get_prompt_versions
    • restore_prompt_version

  • Documents7

    • list_documents
    • get_document
    • create_document
    • update_document
    • delete_document
    • get_document_versions
    • restore_document_version

  • Collections7

    • list_collections
    • get_collection
    • create_collection
    • update_collection
    • delete_collection
    • add_to_collection
    • remove_from_collection

  • Search1

    • find_items

  • ChatGPT App2

    • search
    • fetch

  • Deep Search3

    • deep_search
    • deep_read
    • deep_expand

  • Reasoning1

    • reason

Deep Search

Search inside documents, not just titles.

Most tools stop at the filename. Deep Search lets your client search within document content, land on the exact paragraph, and walk the document like a table of contents without loading all of it.

Vector-powered content search
Finds passages by meaning, not keyword. Ask a question and get the paragraphs that answer it, ranked.
Hierarchical navigation
Move up to the parent section, down into paragraphs, or sideways through siblings. Five directions from any hit.
Session-aware deduplication
Chunks already returned in the same session are filtered out, so a second search surfaces what the first one missed.

deep-search.log

> deep_search "rate limiting strategy"   0.82  API Guidelines › Rate limits › Sliding window  0.74  API Guidelines › Rate limits › Fail-open  0.61  Incident notes › 2026-03 › Cold starts > deep_expand chunk_01 up   API Guidelines › Rate limits  ├─ Sliding window        (current)  ├─ Fail-open             (next)  └─ Per-instance counters (next) > deep_read chunk_01   "Requests are counted in a sliding window per key.   When the store is unreachable the limiter fails   open so a Redis outage never blocks writes."
Deep Search tools reference

Two ways to connect.

Same account, same 29 tools either way.

Diagram of the remote connection. AI chats (ChatGPT, Claude, Grok) sign in with OAuth, and other clients send an API-Key header. Both connect two-way over HTTPS, using Streamable HTTP, to Context Repo, which accepts OAuth sign-in or a scoped API key: same account, same tools. A highlighted two-way path joins AI chats to OAuth.

Remote / Streamable HTTP

Point your client at one URL.

For clients that implement MCP OAuth or send custom authorization headers. No Node.js on your machine. Grok uses the dedicated /grok-mcp endpoint and sees the same tools.

  1. 01Add https://contextrepo.com/mcp as a remote MCP server.
  2. 02Sign in with your Context Repo account when the client asks, or set an Authorization: API-Key header.
  3. 03Ask “List my prompts” to confirm the connection.
Streamable HTTP setup
Diagram of the local connection. On your machine, code editors (Cursor, VS Code, Windsurf) launch the context-repo-mcp bridge through npx and talk to it two-way over stdio. The bridge connects two-way over HTTPS to Context Repo with a scoped API key: same account, same tools. The highlighted path runs from the editors through the bridge to the API key.

Local / stdio bridge

Run the open-source bridge.

For clients that launch a local command. context-repo-mcp v3 forwards to the hosted endpoint, so tools and results stay identical to the remote path.

  1. 01Create an API key under Dashboard › Settings › API Keys with the prompts.read and documents.read scopes.
  2. 02Add npx -y context-repo-mcp with CONTEXTREPO_API_KEY to your client config (Node.js 18+).
  3. 03Restart the client and ask “List my prompts”.
Cursor and Claude Desktop guides

Working in a shell, a script, or CI? The CLI reaches the same prompts, documents, and collections through the REST API.

Get the CLI

Your questions, answered.

  • The Context Repo MCP Server provides 29 tools for prompts, documents, collections, search, reasoning, and hierarchical retrieval. Clients must support authenticated Streamable HTTP or be able to launch the local stdio bridge.

Open-source bridge

Zero copy-paste. Full context.

All 29 tools in every plan.

context-repo-mcp on GitHub

Illustrative activity: documents and prompts lighting up as connected AI tools read and write them.