Note Taker MCP
Allow the model to write scratch notes during reasoning phases for more reliable detail surfacing in long context tasks
Installation
npx note-taker-mcpAsk AI about Note Taker MCP
Powered by Claude · Grounded in docs
I know everything about Note Taker MCP. Ask me about installation, configuration, usage, or troubleshooting.
0/500
Reviews
Documentation
An important pre-read caveat. This project is 100% codex written. I wrote nothing execpt the prompts to execute this. I plan to take a pass at cleaning up the code where needed but as of now the code is rough but working.
Notes MCP
The goal of the notes mcp server is to allow for LLMs to write and read small chunks of information that they think may be important to reference later.
Design
- The notes will be stored as filestore files with a unique uuid for each file.
- A summary of the note (or details the model provides on when it would want to use that note again) will be persisted to a local vector store along with a metadata field on the entry containing the uuid to reference back to the file.
- The vector store is a local Chroma instance backed by an on-device embedding function, persisted under
.notes/chroma/.
The system is designed with four core actions to interact with this system
Actions
Write Note
Description: This should create a new note, persist it to disk, and add an entry into the Vector Store.
Args:
- note: str - the note to persist to disk
- when_to_use - a summary of the note and when the model would want to use this note
- tags: list[str] (optional) - metadata tags to aid filtered search
Search Notes
Description: Returns up to five notes matching the query, optionally filtered by tags and score threshold.
Args:
- query: str - the search query
- tags: list[str] (optional) - only return notes that share at least one of these tags
- score_threshold: float (optional, default 0.75) - minimum similarity score required
- k: int (optional, default 5) - max results to return
Return:
- results: list of objects each containing
note,note_id,score,when_to_use, andtags
Remove Note
Description: Deletes a note from filestore & vectore store
Args:
- note_id: str - the id of the note to delete
Update Note
Description: Updates an existing note. Will overwrite the entire note with the new content provided, is not diff based
Args:
- note_id: str - the id of the note to update
- note: str - the new note content to overwrite the old note
Running the server
Codex config.toml
[mcp_servers.notes]
command = "uvx"
args = ["note-taker-mcp"]
Cursor / Claude Code
"notes": {
"command": "uvx",
"args": [
"note-taker-mcp"
]
}
Running Locally
Requirements: Python 3.10+, uv, and the fastmcp/chromadb dependencies installed via uv sync.
Quick start:
uv sync
uv run notes-mcp
By default, each server process creates its own session-scoped data root under /tmp/notes-<session-id>/, with note bodies in notes/ and the vector index in chroma/. When the server shuts down, this temp directory is deleted. Set NOTES_MCP_DATA_DIR=/custom/path (or pass data_dir to build_server) to use a persistent location instead.
Getting the model to reliably take notes
Reference example.AGENTS.md
Development notes
- Entry point:
notes_mcp/server.py(script name:notes-mcp). - Storage layer:
notes_mcp/storage.pywrites UTF-8 note files keyed by UUID. - Vector index:
notes_mcp/index.pyuses Chroma’s persistent client and default local embedding function. - Tests:
uv run python -m pytest.
The server runs over stdio via uv run notes-mcp and registers four MCP tools: write_note, search_notes, update_note, and remove_note.
