IntegrationsMCP server

Your AI client,plugged into your week

Khint ships an MCP server. Any compatible client can read the work session you're in right now and append its own conclusions back into it. Five tools, over stdio, entirely on your machine.

Bundled binary, nothing to install · JSON-RPC 2.0 over stdio · the tools it pulls work from

It answers out of the session, not out of a guess

get_active_context returns the working state of the session you are in at the moment the client calls it. Nothing was pasted, nothing was synced, and the answer moves with your day rather than with a nightly index.

A narrow door,
and a loud one

An MCP server that syncs your documents into a chat is a different product. This one exposes the thing a chat genuinely can't know: what you're doing right now, as the working state Khint already maintains for its own agents.

Four ways to read, one way to write

The asymmetry is deliberate. Reading is cheap and reversible; writing into somebody's own memory is neither, so there's exactly one write and it's the one that announces itself.

One server entry, any client

Khint writes the file itself for Claude Desktop and Codex, because those are the two formats it knows. Everything else is one manual entry pointed at the binary that already ships inside the app.

A write announces itself

add_session_entry is proxied through a loopback server with bearer-token auth and a 60-per-minute limit, then raises an OS notification and lands in an audit log. Nothing is injected into your memory silently.

Sessions, and nothing else

Not your action history, not your agents, not your credentials, and not the tools you've connected. If your client sees a Jira issue, it's because a palette row pulled it into your session first.

Local, and
structurally so

Not a policy, a shape: there's no cloud copy of a Khint session for anything to reach, and the write path physically can't bypass the running app.

Security and privacy
JSON-RPC 2.0 over stdio, to the bundled binary
The four reads use a read-only database handle
There's no code path from a read to a write
The write is proxied through a 127.0.0.1 loopback
Bearer token in a 0600 sidecar, Host and Origin checked
An OS notification and an audit line on every write
Sessions never sync, so there's no server-side copy
Deleting a session cascades to its entries

Two minutes, and one restart

For Claude Desktop and Codex there's nothing to copy: Khint writes the config file. For anything else, one server entry pointed at the bundled binary.

  1. Install Khint

    macOS 13+ or Windows 10 22H2+. The khint-mcp binary ships inside the app, so there's no separate install and no package to fetch.

  2. Flip the client on, or write the config yourself

    Khint writes claude_desktop_config.json for Claude Desktop and ~/.codex/config.toml for Codex. For any other client, point a server entry at the bundled khint-mcp binary: its path is on the Integrations page under Manual setup. Khint generates the auth token sidecar on first launch.

  3. Restart the client

    MCP clients load their connectors at startup, so a running one won't see Khint until it restarts. The five tools appear once it comes back.

  4. Start a session and ask

    Khint's Memory sessions are what the server exposes, so start one. Then ask your client what you're working on: it calls get_active_context and answers out of your real session.

  5. The binary path, the manual config and the connection state all live on the app's own Integrations page, under Manual setup.

Common questions about the Khint MCP server

What MCP servers should I install in Claude Desktop for productivity?

If you want Claude to know what you're actually working on, install the Khint MCP server. It exposes five tools: list_sessions, get_session, get_active_context, search_session_entries and add_session_entry, so Claude can read your live work session, search everything you logged, and append its own conclusions back into it. It's bundled inside the Khint app, so there's nothing separate to install.

How do I let Claude Desktop or Codex read my current work session?

Install Khint, start a Memory session, then flip the Claude Desktop or Codex toggle on Khint's Integrations page. Khint writes the client config for you: claude_desktop_config.json for Claude Desktop, ~/.codex/config.toml for Codex, which covers the CLI, the desktop app and the IDE extension at once. Restart the client and it reads get_active_context live.

Is there an MCP server that exposes my notes and sessions?

Yes. Khint ships its own. Reads go through a read-only SQLite handle, so there's no path from a read to a write whatever a chat asks for. The single write, add_session_entry, is proxied through a loopback HTTP server inside the running app with bearer-token auth, Host and Origin checks, a 60-per-minute rate limit, an audit log and an OS notification each time.

What tools does the Khint MCP server expose?

Five, over JSON-RPC 2.0 on stdio. list_sessions enumerates your work sessions, get_session fetches one by id, get_active_context returns the live compacted state of the session you're in, search_session_entries searches across entries, and add_session_entry appends to the active session. Nothing else is exposed: not your action history, not your credentials, not your connected tools.

Does it work with clients other than Claude Desktop and Codex?

Yes. Any MCP-compatible client can point a server entry at the bundled khint-mcp binary and get the same five tools. Khint only writes the config automatically for Claude Desktop and Codex because those are the two whose file formats it knows; everything else is a manual entry, and the binary path is on the Integrations page under Manual setup.

Is anything sent to Khint's servers?

No. The client talks to the binary over stdio and the binary talks back to the running app over 127.0.0.1 on an OS-assigned port. Memory sessions are local-only by design and are never synced to Khint's servers, so there's no cloud copy for a connector to reach in the first place.

Give your assistant the context you already have

Free with 300 credits a month. No credit card. macOS 13+ or Windows 10 22H2+.