Your notes, queryable from every AI tool you use

MCP is an open standard that Anthropic published in late 2024. It exists because every AI assistant ends up needing the same integrations: a filesystem, a GitHub account, a database, your notes app. Before MCP, each vendor wired those connections in its own way, none of them compatible. MCP gives everyone the same socket to plug into.

The shape is simple. On one side, an MCP server exposes some data, for example a GitHub repo, a Postgres database, or your Korely vault. On the other side, an MCP client reads from it, for example Claude Desktop, Cursor, Zed, ChatGPT Desktop, n8n, Continue, Warp Terminal, or Raycast. Once an MCP server exists, every compliant client can read it right away. No custom integration code on either side.

Korely is one of those servers. The vault on your disk becomes a knowledge source any compliant AI tool can search, read, and reason on. You set it up once, and from then on the assistants you already use know your notes.

Korely's MCP server follows the "pure data provider" pattern. No LLM call runs on our side, no opinionated synthesis, no hidden prompts. The tools return Markdown, and the assistant on the other end reasons on it. The CLI exposes the exact same five tools, just invoked from a shell instead of from a chat.

• korely_search(query, limit?, expand_with_graph?)

  Hybrid full text plus vector search across the vault. Returns the top matches with snippet and score.

• korely_read_item(item_id)

  Returns the full Markdown content of one note or document.

• korely_get_related(item_id)

  Returns notes connected via shared entities in the knowledge graph. Great for questions such as "what else did I write about this?"

• korely_list_notes(folder?, tag?, limit?, offset?)

  Paginated browse, good for workflows such as "scan all my January meeting notes".

• korely_list_folders()

  Returns the vault's folder structure. An orientation tool, mostly for assistants exploring the vault for the first time.

Pro adds three write tools (save_note, create_task, save_session), so the assistant can persist its work back into your vault when you allow it.

The CLI is the same five read tools as MCP, but invoked from a shell. One command, one answer, ready to pipe into anything else. It's what you reach for when you want to drive the query yourself instead of letting an AI agent wander around looking for it.

There are two modes, and the distinction matters.

• Local (Free). Reads the vault directly from disk, on the same machine where Korely desktop is installed. No network, no auth, just the SQLite file the app already uses. Use this for interactive lookups from your laptop and for scripts that run while you're at the desk.
• Cloud (Pro). Reads from your synced vault in Korely cloud, so you can run the CLI from any machine that's signed in, for example a server, a CI runner, a phone shortcut, or a cron job on a small VPS. Same subcommands, same output, different data source.

Two reasons to keep the CLI around even if you already love MCP.

Token economics. When an AI agent runs over MCP, it usually explores. It searches, then reads, then searches again, sometimes five or ten tool calls per question, each one growing the context window. When you call the CLI yourself, you pull exactly what you need and hand a compact result to the model. Same vault, same data, a fraction of the spend. For instance, asking Claude over MCP "what did we decide about Q3?" might cost around 10k tokens of exploration. The same answer via korely search query="Q3 decisions" | head -3 | claude "summarize" costs roughly 2k.

Composition. The CLI plays nicely with the rest of your terminal world. Pipe it to jq, to grep, to your favourite local LLM (such as Ollama or LM Studio), to a shell script that runs every morning, to a CI job that prepares context before a build. You're not locked into one chat interface, which matters when you want zero token cost workflows or scriptable automation.

A few real examples:

# Find every note that mentions a topic, top 5
korely search query="Q3 pricing decisions" limit=5

# Read one note in full
korely read id=<note-id>

# Find notes connected to one note via the knowledge graph
korely related id=<note-id>

# List everything in a folder, filtered by tag
korely list folder="Meetings" tag="standup"

# Pipe to a local LLM for a zero token cost summary
korely search query="what did the customer say" | ollama run llama3.1 "summarize"

# Cron job, every Monday: open action items as JSON, ready for any pipeline
korely search query="open action items" --format=json

Output is Markdown by default, JSON with --format=json when you want to script. Subcommands follow the same naming as the MCP tools, so anything you learn in one place transfers to the other.

Claude Desktop. Open Settings, go to Developer, and edit claude_desktop_config.json. Add Korely as a stdio MCP server, pointing at the binary that ships inside the Korely app bundle. Restart Claude. The Korely tools appear in the tools dropdown right away.

Cursor. Settings, then Features, then MCP Servers, then Add. Same config shape as Claude Desktop, only it lives in ~/.cursor/mcp.json.

Zed, ChatGPT Desktop, Warp, Raycast. Same pattern, each with their own config location. The exact path is documented inside each client's settings panel.

For n8n, see the Cloud MCP section below. n8n runs in the cloud, so it connects to the Pro cloud endpoint rather than to a local server.

The CLI ships inside the Korely desktop app on macOS, Windows, and Linux. After install, register it on your PATH and you can call it from any terminal on the same machine. No auth needed for the local vault: if you can read your home folder, you can read your notes.

# 1. Install the CLI on your PATH (one time)
korely cli:register

# 2. Verify it's there
korely --version
# Korely CLI 1.0.0

# 3. Try a search against your local vault
korely search query="any topic" limit=3

For cloud mode (Pro), the setup is one extra step: korely auth --cloud. It opens a browser for OAuth, stores a refresh token locally, and from then on every command can talk to your synced vault from any machine. Useful when you want to run the CLI from a server, a CI runner, or any place that isn't your laptop.

The full command reference and editor integration examples (for instance Vim, Emacs, VS Code) ship inside the desktop app's help menu.

Local MCP is great when the assistant runs on the same laptop as Korely. The moment you want to query your vault from a different machine, such as a work computer, a phone, or an n8n cloud worker, local MCP can't help.

Pro Korely syncs the vault to Korely cloud and exposes an OAuth protected MCP endpoint. Add it to Claude on your work machine, give consent in the browser, and the assistant can read your full vault remotely. The endpoint is multi tenant, and every user gets their own OAuth isolated namespace.

For automation workflows (such as n8n, Zapier with MCP, or custom scripts) cloud MCP unlocks a layer that the local only setup simply cannot. In n8n, for example, you add an MCP credential pointing at your Korely cloud endpoint, drop the MCP node into a workflow, and you have things like "every Monday, ask my vault what's open from last week's standups, then post it to Slack" running on their own.

The CLI gets the same upgrade. Run korely auth --cloud once, then any command (for example korely search query="Q3") hits the cloud vault instead of the local SQLite. Same data, different machine, no setup beyond the OAuth handshake.

• Ask Claude "what did the customer say in our last three calls?". Claude fires korely_search and korely_read_item, then returns the synthesised answer with the source notes linked inline.
• Ask Cursor "find the architecture note where I decided on pgvector vs Pinecone". Cursor uses the same MCP tools to pull the note straight into the chat context, no scrolling through your vault needed.
• n8n workflow: trigger every Monday, ask the cloud MCP "what action items are open from last week's standups?", post the result to Slack. No custom integration code, just one MCP node.
• ChatGPT Desktop on macOS: "draft a follow up email to Sarah Chen based on my last meeting note with her". ChatGPT pulls the note, drafts the email, asks if you want to send.
• Terminal one liner, no AI involved at all: korely search query="API rate limit" | head -1. Faster than opening the GUI to look up that one detail you half remember.