One address
Point your assistant at Wend
Wend speaks MCP, so whatever you already use can ask about the people you know, propose what it learns, and draft on your behalf. It reads the same memory Wendy works from, so a fact your assistant confirms is one she can act on. Sign-in happens inside your own client. There is no key to paste, except for the Claude extension, which runs on your machine and asks for one itself.
The server:
https://www.trywend.io/api/mcpStreamable HTTP, OAuth 2.1. Always the www host: the bare domain redirects, and some clients drop the sign-in header across that hop, which looks like a broken server rather than a redirect.
Every client
Pick yours
One click where the client supports one, one command where it does not, and one block to paste where neither is true.
Claude Desktop
Open it and Claude installs it. It asks for a Wend key once, which you mint at https://www.trywend.io/account/agents; it is kept in your keychain.
Cursor
Cursor opens, shows the server, and you approve it. Sign-in happens inside Cursor.
VS Code, from a terminal
Run this.
code --add-mcp '{"name":"wend","type":"http","url":"https://www.trywend.io/api/mcp"}'Claude Code
Run this.
claude mcp add --transport http --scope user wend https://www.trywend.io/api/mcpThen run /mcp in Claude Code and sign in.
Codex
Run this.
codex mcp add wend --url https://www.trywend.io/api/mcpCodex also needs experimental_use_rmcp_client = true at the top of ~/.codex/config.toml, or it ignores remote servers entirely.
Gemini CLI
Add this, in ~/.gemini/settings.json.
{
"mcpServers": {
"wend": {
"httpUrl": "https://www.trywend.io/api/mcp",
"oauth": {
"enabled": true
}
}
}
}Gemini reads httpUrl for streamable HTTP. The url key means something else and connects to nothing.
Anything else
Add this, in wherever that client keeps its MCP servers.
{
"mcpServers": {
"wend": {
"type": "http",
"url": "https://www.trywend.io/api/mcp"
}
}
}Most clients take this shape. If yours wants a transport name, it is streamable HTTP.
Unlimited through your own AI. The reasoning runs on the plan you already pay for; Wend serves the memory.
Want the longer version, with each client's own flow and the trap it hides? Connect Wend to your agent walks Claude Desktop, Claude Code, Cursor, VS Code, Codex and Gemini step by step. Working out of a terminal? Wend from the command line covers the three CLIs and why the reasoning runs on the subscription you already pay for.
Schedule it
Three prompts worth scheduling
Your assistant already runs on a schedule. Give it one of these and it runs on your plan; Wend answers from your Mac when it is on and from your cloud copy when it is not, and anything that would leave for another person comes back as an approval link.
Morning brief
Every day at 8:00. Your brief, who is waiting on a reply, today's birthdays, the replies drafted, and the approval links.
Ask Wend for my brief, my reply debt, and today's birthdays; draft the replies; give me the approval links.Sunday Six
Sundays at 19:00. Six people you have not talked to this week, why each one, and Wendy's words for each.
Run my Sunday Six with Wend: six people I have not talked to this week, why each one, and Wendy's words for each. Give me the approval links.Pre-meeting brief
Every morning, or ten minutes before a meeting. For each meeting on the calendar: who is on it, the last contact, what you owe them, one opener.
Brief me with Wend on each meeting on my calendar today: who is on it, when I last spoke to them, anything I owe them, and one thing to open with.Claude also reads a skill file that teaches it Wend's vocabulary and rituals: the Sunday Six, who you are drifting from, the four send modes, and what an approval link is. The Wendy skill is one file; point Claude at it or add it to your project.
What it can do
Read, propose, draft, send
Four permissions, granted separately. An assistant you want reading your graph and never sending anything gets exactly that, and the server enforces it on every call.
Reading answers who you know, how you met, and who you are overdue to talk to. Proposing puts a fact in the review queue with the assistant's name on it, and nothing reaches your graph until you confirm it.
They are the four scopes on the token: wend.read, wend.propose, wend.act.draft and wend.act.send. A client only sees the tools its scopes allow, so a read-scoped assistant does not get a send tool it could be talked into calling, and a send asked for anyway comes back as an error naming the scope it would have needed. The check is on the credential, never on a line in a tool description.
Drafting and sending are separate on purpose. Wendy writes the exact words, you read them, and only then does anything go out. Nothing sends as you unattended unless you switch on an Autopilot lane for one person and one kind of message. Four modes for each person: she drafts and you send; you approve the exact words and it goes out from your own Messages; she sends as herself, signed; or your Wendy talks to theirs.
A key for the Claude extension is minted at your account settings. Every other client signs in over OAuth and holds no key at all.
The memory itself lives on your Mac
Wend reads what your machine already holds, keeps it in one place, and answers your assistant from there. Wendy works from the same memory and reaches you on Telegram and by email, so an assistant is a way in rather than the only one.