One URL for ChatGPT, Claude, Gemini and Perplexity. One line for Codex, Claude Code, Cursor and VS Code.

Claude as a custom connector

1. In claude.ai or Claude Desktop open Settings → Connectors → Add custom connector. The Connect page of the Agentu web app opens this dialog with the values prefilled.
2. Name it Agentu Memory and paste https://memory.agentu.com/mcp, then Add and Connect.
3. Claude opens Agentu's consent page. Paste a key — ideally a narrow one such as read:Work, write:Work — and choose Allow.
4. In a conversation, enable the connector. Claude now has seven tools: recall, remember, remember_many, remember_file, find_file, list_categories and list_projects (what each does is in the Any MCP client tab).

Behind the scenes this is OAuth 2.1 with PKCE: Claude never sees your key, only a token bound to it. Revoke the key and the token dies with it.

Claude Desktop and Claude Code through the Agentu Bridge

The local option is the Agentu Bridge on your computer: it runs as a local MCP server over stdio (agentu-bridge mcp), decrypts on this computer and hands the app only the few memories that match a question. The keys stay with you. There is no npm package to install. In build The packaged Bridge for Mac is in build; until it ships, the Bridge is built from the repository.

One click. agentu-bridge connect claude-desktop writes the entry below into claude_desktop_config.json (~/Library/Application Support/Claude/ on Mac, %APPDATA%\Claude\ on Windows, ~/.config/Claude/ on Linux); then restart Claude Desktop. Manual: Settings → Developer → Edit Config, paste the entry with your real paths (which node and the Bridge's dist/agentu-bridge.mjs), save, restart.

{
  "mcpServers": {
    "agentu": {
      "command": "/path/to/node",
      "args": ["/path/to/agentu-bridge.mjs", "mcp"]
    }
  }
}

Claude Code. agentu-bridge connect claude-code runs this for you (--scope user so it applies in every folder). Manual: the same command; claude mcp list shows it, claude mcp remove --scope user agentu removes it.

claude mcp add --scope user --transport stdio agentu -- /path/to/node /path/to/agentu-bridge.mjs mcp

Prefer the hosted link instead of the Bridge? Claude Desktop takes the custom connector exactly as claude.ai does (first tab); for Claude Code use the --transport http line in the next tab. The Bridge only ever writes the agentu entry in each file, keeps a .bak-<timestamp> copy next to it, and agentu-bridge connect --remove <app> takes the entry out again.

ChatGPT connectors

1. Open Settings → Connectors → Create (custom connectors require developer mode on Plus, Pro, Business or Enterprise plans).
2. Name: Agentu Memory. URL: https://memory.agentu.com/mcp. Authentication: OAuth.
3. ChatGPT opens Agentu's consent page. Paste a scoped key and allow.
4. Enable the connector in a chat. Ask ChatGPT to remember something, then ask Claude about it — that is the whole point.

Codex, Claude Code, Cursor and VS Code

These apps sign in through a loopback callback on your computer. In build Agentu's sign-in for loopback callbacks is being fixed first, with a regression test in CI; each app is marked live here when the connection, the tool list, a read and a write all pass.

• Codex (CLI, desktop app and IDE extension share one file, ~/.codex/config.toml; agentu-bridge connect codex writes it for you), then codex mcp login agentu opens the Agentu approval screen. In the Codex app: Settings → MCP servers → Add server, name agentu, URL https://memory.agentu.com/mcp.

codex mcp add agentu --url https://memory.agentu.com/mcp

• Claude Code with the hosted link (--scope user applies it in every folder; /mcp inside Claude Code signs in). The local path through the Bridge is in the previous tab.

claude mcp add --scope user --transport http agentu https://memory.agentu.com/mcp

• Cursor: open the install link below, or Settings → Tools & MCP → Add Custom MCP and paste { "url": "https://memory.agentu.com/mcp" } as the agentu entry in ~/.cursor/mcp.json. agentu-bridge connect cursor does the same and prints the link.

cursor://anysphere.cursor-deeplink/mcp/install?name=agentu&config=eyJ1cmwiOiJodHRwczovL21lbW9yeS5hZ2VudHUuY29tL21jcCJ9

• VS Code: open the install link below, or Command Palette → MCP: Add Server → HTTP → https://memory.agentu.com/mcp → name agentu. agentu-bridge connect vscode writes the entry into your user mcp.json and prints the link.

vscode:mcp/install?%7B%22name%22%3A%22agentu%22%2C%22type%22%3A%22http%22%2C%22url%22%3A%22https%3A%2F%2Fmemory.agentu.com%2Fmcp%22%7D

OpenAI Responses API (remote MCP tool)

For your own agents and products. Give the tool a key scoped to exactly the categories it needs.

{
  "model": "gpt-5",
  "tools": [{
    "type": "mcp",
    "server_label": "agentu_memory",
    "server_url": "https://memory.agentu.com/mcp",
    "headers": { "Authorization": "Bearer agk_live_…" },
    "require_approval": "never"
  }],
  "input": "What did I decide about the board meeting?"
}

Perplexity custom connector

1. On a Pro, Max or Enterprise plan open Settings → Connectors → Add custom connector.
2. Name: Agentu Memory. URL: https://memory.agentu.com/mcp. Authentication: OAuth (or an API key if your plan offers it: paste a scoped key).
3. Approve on Agentu's consent page. Enterprise admins can share the connector organisation-wide.

Gemini app and Gemini Enterprise

1. Gemini app: Settings → Connected apps → Add custom app and paste https://memory.agentu.com/mcp; authorise on Agentu's consent page. At launch Google limits custom apps to personal Google accounts, United States, English, 18+, with the activity setting on.
2. Gemini Enterprise: an administrator adds a custom MCP server (streamable HTTP) with the same URL and OAuth; the organisation policy must allow custom servers.
3. Gemini API: pass https://memory.agentu.com/mcp as an mcp_server tool with your key in the headers.

Any MCP client

Remote: Streamable HTTP at https://memory.agentu.com/mcp with OAuth 2.1 (discovery documents are published; Authorization: Bearer <key> for your own agents). Local: the Agentu Bridge over stdio, agentu-bridge mcp, with the keys on your computer. The server announces itself as Agentu Memory and exposes seven tools; every one takes an optional project (a name or id from list_projects), and without it uses the project this connection was given.

Only the categories the user allowed for this app are visible, and only connected projects are reachable; a reviewer's project is read-only. Every read and write is written to the user's activity log.