funes-vault-app

<h1 align="center">Funes Vault</h1>
<p align="center"><b>Give AI agents your memory. Not all of it.</b><br>
A free, open-source Mac menu bar app that keeps your personal context in a private vault on your Mac, and shares it with Claude, Cursor, Codex and other AI agents only as you allow.</p>

<!--
Badges, once the listings exist:
<p align="center">
  <a href="https://github.com/bereciartua/funes-vault-app/releases/latest"><img alt="Latest release" src="https://img.shields.io/github/v/release/bereciartua/funes-vault-app?label=release"></a>
  <img alt="macOS 15 or later" src="https://img.shields.io/badge/macOS-15%2B-black">
  <a href="LICENSE"><img alt="License: Apache-2.0" src="https://img.shields.io/github/license/bereciartua/funes-vault-app"></a>
  MCP Registry (com.funesvault/funes), mcpservers.org and glama.ai badges go here.
</p>
-->

<p align="center">
  <a href="https://github.com/bereciartua/funes-vault-app/releases/latest/download/FunesVault.dmg"><b>Download for Mac</b></a> ·
  <a href="#get-started">Get started</a> ·
  <a href="docs/MCP.md">Agent setups</a> ·
  <a href="CHANGELOG.md">Changelog</a>
</p>

<p align="center">
<picture>
  <source media="(prefers-color-scheme: dark)" srcset="docs/images/vault-memory-dark.png">
  <img alt="The Funes Vault window: the Vault list with search and filters, and one memory open with its text, categories, sensitivity level and how often it was shared." src="docs/images/vault-memory-light.png" width="720">
</picture>
</p>

- **Each agent gets only what you choose.** Its own key, and the access you set: read, suggest or save; up to a sensitivity level; in the categories you pick. A new agent can start with nothing.
- **You approve the exact text.** Before a memory goes to an agent, a small panel shows it, and you can untick what shouldn't go. An approval is used once, within 15 minutes, and is cancelled if anything changes.
- **Agents suggest; you decide.** What an agent wants to remember waits in your Inbox, where you edit it, apply it or reject it.
- **Everything stays on your Mac.** No account, no server of ours, no cloud copy. Search runs on a small model built into the app. Cloud agents such as claude.ai and ChatGPT can reach it only through a tunnel you run, with Remote Access (experimental, off by default).

**Status:** public preview (v0.1), with best-effort support. **Works with** Claude Code, Claude Desktop, Cursor and Codex, with ready-made setups for VS Code (Copilot), Gemini CLI, Zed, Cline, JetBrains AI Assistant, Devin Desktop and any other MCP client. **From the cloud (experimental):** claude.ai, ChatGPT, Gemini Enterprise, the Anthropic and OpenAI APIs, and cloud coding agents, through Remote Access.

## Get started

1. **Install.** Download [`FunesVault.dmg`](https://github.com/bereciartua/funes-vault-app/releases/latest/download/FunesVault.dmg), open it and drag Funes Vault to **Applications**. If you open it from the disk image or Downloads instead, it offers to move itself there. Or use [Homebrew](https://brew.sh):

   ```sh
   brew install --cask bereciartua/tap/funes-vault
   ```

   Funes Vault needs **macOS 15 or later**, on Apple silicon or Intel.

2. **Open the app.** It lives in the menu bar. The Vault opens with a short setup checklist:
   1. **Seed your vault.** Answer a few optional questions, import your ChatGPT memories or paste Claude's, or import a Markdown file. Everything goes to the Inbox first, and anything that looks like a secret is never imported.
   2. **Review your Inbox.** Apply what's right; edit or reject the rest.
   3. **Add an agent.** Pick its tile (agents found on your Mac are marked) and a starting access. The default, **Read low-sensitivity memories, ask first**, lets it read memories up to Low and asks you before each share.
   4. **Connect it.** **Add to Claude Code…** (or Claude Desktop, Cursor or Codex) shows exactly what it will change or run first, backs up a settings file before changing it, and never writes the key into the agent's settings. For other agents, copy the setup from the agent's **Connect** tab. The step is done when the agent's first request arrives.

3. **Ask your agent.** For example: *"Plan a week of dinners for me"* or *"Set up this TypeScript project the way I like it."* The agent asks Funes for context, you see the panel, and you choose what to share.

You can also capture a thought at any time with **⌥⌘M**, **Remember…** in the menu, or **Services ▸ Remember in Funes** in any app.

Setups for every agent, the tool reference and troubleshooting are in the [MCP guide](docs/MCP.md).

### From cloud agents (experimental)

claude.ai, ChatGPT and cloud coding agents run on their vendor's servers, so they can't reach an app on your Mac directly. **Remote Access** lets them, through a tunnel you run, such as [Tailscale Funnel](https://tailscale.com/kb/1223/funnel):

1. Open **Remote Access** in the sidebar and follow its four steps: choose a tunnel, turn Remote Access on and start the tunnel, paste its address, test. Funes never starts a tunnel; it shows the command.
2. Add the agent with a cloud tile in **Add Agent…** (claude.ai, ChatGPT, …), or turn on **Allow cloud access** on an agent's **Cloud** tab.
3. Connect it: apps that sign in (claude.ai, ChatGPT, Gemini Enterprise) use **Connect a Cloud App…** and a six-digit code you compare on the Mac; the others use the agent's remote token. The agent's **Cloud** tab shows the steps for each.

<picture>
  <source media="(prefers-color-scheme: dark)" srcset="docs/images/remote-access-dark.png">
  <img alt="The Remote Access page, on and marked Experimental: the tunnel is reachable and running (Tailscale Funnel, forwarding to 47627), the public address and the MCP URL with its secret path, the port, when to turn off, and the agents with cloud access." src="docs/images/remote-access-light.png" width="720">
</picture>

A cloud agent gets exactly the access you set for it, and the same panels on your Mac: if you're away, its requests wait and expire as for any agent. While Remote Access is on, the vault is reachable through a public address, and what you share with a cloud agent goes to its vendor's servers. It's off by default, and **Turn Off Remote Access** in the menu bar cuts every cloud agent off at once. The [MCP guide](docs/MCP.md#use-from-cloud-agents) has the details.

## How you stay in control

Funes is the only thing that holds your memories. Agents talk to it, never to the vault directly, and it checks every request against what you allowed.

### You choose what each agent may see

Every agent has its own access: whether it may **Read**, **Suggest** or **Write**; a **maximum sensitivity** (Public, Low, Internal, Sensitive, Restricted or Secret); the categories it may or may not use (a denied category always wins); whether Funes **asks before sharing** and **asks before saving**; and an optional expiry date. Health, Finance and Legal memories are Sensitive by default.

<picture>
  <source media="(prefers-color-scheme: dark)" srcset="docs/images/agents-access-dark.png">
  <img alt="The Agents page: each agent with its status and a summary of its access, and Claude Code's Access tab open, with switches for Read, Suggest and Write, a slider for the maximum sensitivity, and Allow or Deny for each category." src="docs/images/agents-access-light.png" width="640">
</picture>

### You approve the exact text

When an agent asks for context, Funes finds the memories that match its task and are allowed for that agent. If asking is on, a panel shows each one with a checkbox, what the agent says it's for, how long it keeps what it gets, and how many more matched but can't go to it. **Share** sends only what you ticked; **Deny** sends nothing. The agent waits up to 45 seconds. If you don't answer in time, the request waits in **Pending requests** for up to 15 minutes, and the agent picks up your answer later.

<picture>
  <source media="(prefers-color-scheme: dark)" srcset="docs/images/share-panel-dark.png">
  <img alt="The Share panel: an agent wants to use your memories for a task, with the keywords it is looking for, its stated purpose, a checkbox per memory with its sensitivity, a countdown, and Deny and Share buttons." src="docs/images/share-panel-light.png" width="400">
</picture>

Nothing an agent writes in its task or purpose changes what it may see. Those words only help find relevant memories.

### Agents suggest, you decide

Agents propose memories, and by default each one waits in the **Inbox** as a card you can edit before **Apply** (**Replace** or **Keep Both** for a correction), **Reject** or **Later**. Cards say who proposed what, quote the agent's reasons as reported, point out similar memories, and show a correction as old and new text. Your own captures and imports land there too.

<picture>
  <source media="(prefers-color-scheme: dark)" srcset="docs/images/inbox-dark.png">
  <img alt="The Inbox: a correction from Claude Code showing the old text struck out and the new text, with editable title, text, kind, sensitivity and categories, the agent's reason as reported, and Later, Reject, Keep Both and Replace buttons." src="docs/images/inbox-light.png" width="720">
</picture>

### You see everything, and stop anything

**Activity** lists what each agent asked for and what happened, by day, with filters. Funes stores only ids and reason codes, never text: it shows the memories involved by their current title when you open a row, says what was withheld and why, and links to fix a refusal. **Pause** an agent to refuse its requests while keeping its setup, **Remove…** it to cut it off, or **Pause Funes Vault** from the menu bar for an hour, until tomorrow or until you resume.

<picture>
  <source media="(prefers-color-scheme: dark)" srcset="docs/images/activity-dark.png">
  <img alt="Activity: filters, two pending requests with Review… buttons, and today's rows with the time, who acted, the event and the result, such as Done, Waiting for y