Connect an agent
Any MCP-compatible agent can read your workspaces and, with your permission, write to them. Claude and ChatGPT list Everythings in their directories, so connecting there takes a click and a sign-in. Every other agent takes one address.
Pick your agent
Each page walks through one agent from install to first prompt.
The one address
Every agent connects to the same MCP server URL. The server speaks streamable HTTP and signs you in with OAuth, so in the apps there is no token to paste. In Claude and ChatGPT you can skip even the address and connect from the listing.
https://www.everythings.app/api/mcp
- Claude: claude.ai/directory/everythings, then Connect.
- ChatGPT and Codex: the Everythings plugin, then Install plugin.
- Gemini, Grok, other MCP clients: paste the address as a custom app or server and follow the sign-in.
- CLIs and scheduled runs: the same address with an access token in the Authorization header.
Read-only or read & write
An agent asks for read access, or for read and write access, when it connects. If you are signed out, a page names the agent and asks you to sign in first; then the approval page lists what it asked for, and you Approve or Deny. For an access token you pick the scope yourself when you create it; start with read-only unless the agent has a job that needs writing.
Read only
The agent can search and read your workspaces, things, comments and marks, see what changed since its last run, and read the answers to questions it asked. Nothing it does changes your data.
Read & write
Adds creating, editing, moving, reordering, copying, deleting (to Trash) and restoring things, creating and renaming workspaces, comments and marks, asking you questions, and claiming a thing while it works on it.
A delete over MCP moves a thing to Trash, where you can restore it. No tool empties Trash, deletes a thing for good, deletes a workspace, or invites anyone. Those stay yours, in the app.
What you see afterwards
Everything an agent does is visible, labeled and reversible.
Its name on its work
A thing an agent created or last changed shows “✦ Claude” (or whichever agent it was) at the foot of its page, its tile carries a ✦, a comment it left reads “via” that agent, and a mark it left says “agent” after your name in the list of who marked. When you edit the thing yourself, the last-changed label clears; a thing the agent created keeps its ✦.
Its questions in your inbox
An agent that needs a decision asks on the thing and stops. The question shows as ❓ on the thing and in the 📥 Inbox tile, the app notifies you, and you answer on the web, on the phone or in the chat. Questions, inbox & @Agent.
Your workspace, drawn in the chat
In Claude and in ChatGPT on the web, asking to see a workspace or a thing draws the Everythings view inline: tiles, Markdown, marks and comments, with tiles you can click. The view in the chat.
Notifications when your agent acts
Two notification kinds exist just for agents: your agent asking you a question, and your agent creating or editing a thing. Both are on by default and can be turned off per workspace. Notifications.
Staying in control
Settings › MCP Access Tokens lists every connected agent and every token, each with a revoke button.
- To cut an agent off, revoke its row here. That ends every token it holds for you within a minute, refresh tokens included, so an agent signed in with OAuth stays out until you sign in to it again.
- An agent reaches every workspace you can open, shared ones included. Members of a shared workspace see an agent's writes labeled with its name, and their notifications for its new things and edits end “via” that agent.
- A workspace can carry its own Agent Instructions, written by the owner and read by agents before they write there. Workspaces.
- Agents are available on every plan. A guest in the sandbox cannot connect one until they sign up.
Going further
Once an agent is connected, these pages explain what it can do for you and how to work with it day to day.
Starting a workspace from the Agent workflow template gives an agent task lists, plans and an archive to work in, and you a loop to review it from your phone. The Agent workflow.
Tool names such as get_thing_view appear on the Tools reference, which lists every tool with what it does and which ones ask before they run.
Questions
What is MCP?
The Model Context Protocol is an open standard that lets AI agents use outside tools and data. Everythings runs an MCP server, so any agent that speaks it can search and read your workspaces and, with your permission, write to them.
Do I need an access token?
Only for a run where nobody can sign in, such as a CLI on a server or a scheduled routine. Claude, ChatGPT, Gemini, the Claude Code plugin and Slackbot all sign in with OAuth: you approve access on an Everythings page and copy nothing. See Access tokens & CLIs.
Which agents are supported?
Any client that speaks MCP over streamable HTTP. These pages cover Claude (web, Desktop, Cowork and Claude Code), ChatGPT and Codex, Gemini (Spark apps and the CLI), Grok and Slackbot. The same address works in other MCP clients with the generic configuration on the access tokens page.
Is this the same as the AI Provider key in Settings?
No. That key powers AI Mode and voice inside Everythings, which call OpenAI or Anthropic on your behalf. Connecting an agent works the other way: an outside agent reaches into your Everythings data. The two are unrelated, and either works without the other.
Can an agent see workspaces other people shared with me?
Yes. An agent connected to your account reaches every workspace you can open, your own and the shared ones, and nothing beyond them. The access you approved limits what it can do there, never where.
What name does an agent's work carry?
The name the agent registered when it signed in with OAuth (Claude, ChatGPT, Claude Code and so on), or the name you gave an access token. That label shows on every thing, comment and mark the agent writes, so give each agent its own token when you use tokens.
Can I revoke access later?
Yes. Revoke its row under Settings › MCP Access Tokens. For an agent that signed in with OAuth, that ends every token it holds for you, refresh tokens included, so it is disconnected within a minute; removing Everythings from its connectors or plugins list afterwards only tidies that list. An access token you created yourself stops for good when you revoke it.