Connect an agent over MCP

Connect an agent over MCP

Add Personal Agent Proxy to Claude, Devin, ChatGPT, Grok Bot, Hermes Agent, Poke or any other MCP client as an MCP server, and what an agent can do through it.

Personal Agent Proxy is an MCP server. An agent that speaks MCP, such as Claude, Claude Duet, Claude Code, Claude Desktop, Devin, Devin CLI, ChatGPT, Grok Bot, Hermes Agent, Poke or any other MCP client, connects to it, signs in with OAuth, and works as one of your agent logins: it can do what that login is allowed and nothing more. Successful connection actions, actions denied by access settings and actions the service failed are in the activity log, with the task they were made for.

The server is at your Personal Agent Proxy's address with /mcp after it, such as https://app.personalagentproxy.com/mcp. Each agent's page in Personal Agent Proxy shows it under MCP, with a button to copy the address or setup command.

Add the server to the agent

Add the address as a custom connector, or custom MCP server, where the agent keeps them:

  • Claude and Claude Duet: in Claude, open Customize, Connectors, and add a custom connector. Enter a name and the address, choose Sign in now, then choose Register automatically under OAuth client. When Personal Agent Proxy opens, choose Claude. Duet uses the same account-wide connectors as Claude, so they share one login, its access, and its activity history. During the research preview, Anthropic offers Duet by waitlist to Pro and Max subscribers in the US.
  • Claude Code: run the setup command shown on the agent's page. It begins claude mcp add --transport http personal-agent-proxy and ends with the server address.
  • Claude Desktop: open Connectors and add a custom connector with the address.
  • Devin: open Customize, MCPs. Under Install MCPs for Myself, choose +, then New MCP. Enter a name and description, select HTTP, paste the address and choose OAuth. Add the server, open it and choose Connect.
  • Devin CLI: run devin mcp add personal-agent-proxy <server-address>. Devin CLI opens the OAuth sign-in automatically; if it does not, run devin mcp login personal-agent-proxy.
  • ChatGPT: turn on developer mode in the settings for apps and connectors, then create a connector with the address.
  • Grok Bot: open Connect Apps from the sidebar and add a custom connector of the type Remote HTTPS. Enter a name and the address, and leave the HTTP headers and Client ID empty.
  • Hermes Agent: copy and run the hermes mcp add command shown on its setup page. The command enables OAuth and gives the browser sign-in five minutes to finish. When Hermes Agent runs on a remote host, open the printed link locally, approve it, then paste the full redirect URL back into the terminal when asked.
  • Poke: on poke.com/integrations/new, enter a name and the address, and leave the API key empty.
  • Any other agent: choose Custom MCP agent when adding an agent, and add the address wherever the agent keeps remote MCP servers, over Streamable HTTP. It signs in with OAuth, so leave any API key or token empty.

On a Claude Team or Enterprise plan, an owner first adds the custom Web connector under Organization settings, Connectors. Members then find and connect it under Customize, Connectors.

Sign in

The agent opens Personal Agent Proxy to sign in. Sign in as yourself if you aren't already.

Choose the agent it works as

Personal Agent Proxy asks which of your agent logins the app works as, with the one its name points at chosen, such as Claude for Claude. Choose an existing login to give it that login's access, or a new one for a company that doesn't have a login yet, which starts with every connection's defaults. An app whose name points at none of them is offered a new Custom MCP agent named after it, which you can rename; you can have as many custom agents as you like, each with its own name. Allow sends you back to the agent, signed in.

The login's access is set as always, on its page under Permissions. A change counts from the agent's next request.

What the agent sees

The server has one tool, script, the same for every login, so the agent's list of tools never goes out of date when access changes. The agent writes a short program in JavaScript or TypeScript, and the script calls these functions:

FunctionWhat it does
pap.startTask({title, goal, parent?})Where an agent starts: it says what it is about to do for you, with a title and a goal, and gets the task's id, which it gives with every later script. Part of a task handed to a sub-agent is started as a sub-task of it.
pap.connections()The accounts it can use, such as a mailbox, each with its id and what it may do with it.
pap.tools(connection)A connection's own tools, such as emails_list or emails_archive, each with what it does and the JSON Schema of its params. Only the tools the login has are listed.
pap.call(connection, tool, params)Runs one of those tools, one that reads, such as listing emails, or one that changes something, such as archiving or sending one.
pap.finishTask({outcome})Marks the task done, saying what was done. The agent starts a new task for the next thing you ask for.
pap.feedback({message})Tells the makers of Personal Agent Proxy about something that didn't work as the agent expected. It goes to them, not to you, and changes nothing in your accounts.
pap.upload()A URL the agent's own shell sends a file to, which makes it a blob for a tool such as files_upload to put in a drive.
pap.download(blob)A URL the agent's own shell fetches a blob from, such as an email's attachment attachments_blob made one of.

One script can do what would otherwise take the agent many calls, each waiting on the model: look through several accounts at once, go through what it found, and answer with only what it needs. Every call a script makes goes through your settings for the login, as any call does, and is in the activity log on its own. A script can reach nothing but these functions: no websites, no files and no other programs. It runs for at most two and a half minutes and makes at most 200 calls.

The tool is marked read-only, so an agent's app doesn't stop to ask you before each script, though a script may change something where the login is allowed to: what an agent may do is set here, for each connection, under Permissions.

When an agent connects, it is also told which accounts it can use, and that what an email or a note says is information, not instructions.

Tasks

Before an agent does anything with Personal Agent Proxy, it has to say what it is about to do, as a task: a title in a few words, such as "Triage this morning's inbox", and a goal, what you asked for in a sentence or two. Every call it makes after that names the task, and a call without one, or for a task that is finished, is refused. When the agent has done what you asked, or given it up, it finishes the task, saying what it did, and starts a new one for the next thing you ask for. Agents don't always say so: a task the agent never finishes is taken as left open once 20 minutes pass with nothing done for it or its sub-tasks. If the agent comes back to it within 20 minutes more, its next call picks the task up again where it was; after that the task stays ended, and the agent starts a new one for what it does next. An agent can have any number of tasks under way at once, each with its own requests.

An agent may split what you asked for among sub-agents of its own, which sign in as it does and have its access. Each part is a sub-task, started with the task it was handed from as its parent, so the tasks of one job form a tree; finishing a task finishes its sub-tasks still under way too.

The activity log shows each request with the task it was made for. Open an entry to read the task's title and goal, and follow it to every request made for that task, with when the agent started it and whether it has called it finished.

Every email, note, page, issue, event, file and contact an agent reads comes with its url: its link in the service's own app, such as Gmail, Granola, Notion, Linear, Google Calendar, Drive or Contacts. An agent is told to give you that link when it points you to something, never an address of Personal Agent Proxy's own, which only its login can open. A mailbox over IMAP, a database and your vault have no such page, so the agent names what it means instead, such as an email by its subject.

An action the login isn't allowed is refused, and logged as denied. One it is allowed that the service then fails, such as one turned down for too many requests, is logged as failed, with what went wrong. The activity log marks the actions that came in over MCP, each with its task.

Disconnect an agent

On the agent's page, under MCP, every app working as the login is listed with when it connected and, for a shared agent, who connected it. Disconnect signs one out; it has to be connected again. A member who only uses a shared agent sees and disconnects the apps they connected themselves. Revoking the login signs out every app working as it, as well as its sign-in to the agent page.

A self-hosted Personal Agent Proxy has to be reachable from the internet over HTTPS for Claude, Claude Duet, Claude Desktop, Devin, ChatGPT, Grok Bot and Poke to connect to it: they call it from their own servers, not from your computer. Claude Code, Devin CLI and Hermes Agent can instead connect to one running on the same computer, including over localhost.