Claude.ai and ChatGPT connect to Graft with one URL and no token. Paste https://app.getgraft.ai/mcp (or your self-hosted origin plus /mcp) as a custom connector, and the product discovers Graft’s own sign-in from the endpoint, registers itself, and sends you to the console. There you see which product is asking and you create the agent the connection will be, named for the product, with the connections it may use. From then on every call the product makes is that agent’s: its scope, its working set, its approvals, its history. This is the same Graft as Hermes and OpenClaw connect to. The difference is how the product proves who it is: a chat product has a browser to send you to and no config file to hold a token, so it uses OAuth; a harness on a server has a config file and no browser, so it carries a static token. Both end at one agent each.

​
What you need

  • A Graft account. If you do not have one, the consent step offers Create an account and brings you straight back.
  • On Claude.ai, a plan with custom connectors: Pro, Max, Team or Enterprise at the time of writing. On a Team or Enterprise plan, an owner adds the connector for the organisation.
  • On ChatGPT, a plan with connectors and Developer mode turned on: Plus, Pro, Team, Enterprise or Edu at the time of writing. In a workspace, an admin has to allow connectors first.
  • The URL: https://app.getgraft.ai/mcp for Graft Cloud, or your deployment’s public origin plus /mcp. It has to be https unless it is localhost; both products refuse a plain http server elsewhere.

​
Claude.ai

1

Add the connector

Open Customize → Connectors and choose + → Add custom connector (older interfaces keep Connectors under Settings). Give it a name (Graft) and paste the URL into Remote MCP server URL. Leave the OAuth client id and secret under Advanced settings empty: Graft registers Claude as a client itself, so there is nothing to fill in. Choose Add.
2

Connect

On the new connector, choose Connect. Claude opens Graft in a new window. Sign in if you are not already, or create an account; either way you land on the consent page.
3

Consent, and name the agent

The page reads Connect Claude to Graft and shows where Claude will be sent back (claude.ai). Under Connect as, keep A new agent, keep or change the Name, and under Scope tick the connections this agent may use. You can change both later on the agent’s page. Or pick an agent you already have from the same list; its scope, working set and approvals then apply as they are. Choose Connect. The window returns to Claude and the connector shows as connected.
4

Use it in a chat

In a conversation, open the tools menu (the + or Search and tools control under the message box) and turn the Graft connector on. Graft’s meta-tools appear beside Claude’s own; an authored tool appears as <vendor>__<name> once it is promoted. A tool promoted during a conversation is not offered until the next one; see Tools mid-conversation. When a tool needs your approval, Graft asks once per agent, as a card in the chat: the build approval, a keyless connection, and a connection another agent of yours already holds are answered on the card, and a card you dismiss falls back to a link into the console, where you answer under Pending actions. A secret is never asked for in the chat; the card sends you to the console for it.
Claude’s redirect address is one it registers with Graft when it connects, and the consent page shows its host, so nothing else can claim to be Claude.

​
ChatGPT

1

Turn on Developer mode

Open Settings → Connectors, then Advanced settings, and turn on Developer mode. This is what lets you add a connector that is not in OpenAI’s directory. Team, Enterprise and Edu workspaces need an admin to have allowed it.
2

Create the connector

Back on Connectors, choose Create. Give it a name (Graft) and a short description, paste the URL into MCP Server URL, and under Authentication choose OAuth. Leave the client id and secret empty for the same reason as on Claude. Tick the acknowledgement that you trust the server and choose Create.
3

Consent, and name the agent

ChatGPT sends you to Graft. Sign in or create an account, and the consent page reads Connect ChatGPT to Graft, showing chatgpt.com as where you will be sent back. Keep A new agent or pick an existing one, set the name and the scope, and choose Connect. You return to ChatGPT with the connector connected.
4

Use it in a chat

In a conversation, open the tools menu under the message box, choose Developer mode, and select the Graft connector. ChatGPT asks before it runs a write tool; Graft asks too, once per agent, as the same card in the chat, with the console link as the fallback.
Menu names and plan requirements are the products’ own and move; when one has, the GitHub issues are where to say.

​
Tools mid-conversation

Claude.ai reads Graft’s tool list when a conversation starts and does not pick up tools/list_changed in the middle of one. A tool that acquire promotes during a chat, and the execute__<connection id> tool of a connection made during a chat, are offered from the next conversation on. In the current conversation the agent reaches them through run_tool, the meta-tool that runs any tool in the toolbox by vendor and name. acquire_status reports the new tool’s name when the job succeeds and says to call it through run_tool until the list has it, so nothing is lost by staying in the chat; the next conversation shows the tool first-class, with its own input schema.

​
What the product does, and what Graft does

The product never sees a Graft token you typed. It asks the endpoint without one, is refused with a header that names Graft’s metadata, reads that metadata to find the sign-in, registers itself, and opens the consent page in your browser with a one-time challenge. When you connect, Graft hands the product a code, the product trades it for an access token and a refresh token, and every later request carries the access token. The access token lasts an hour and the product renews it with the refresh token on its own, for as long as the agent lives; each renewal retires the old refresh token, so a stolen one is worth nothing once the product has moved on. None of this needs you after the consent. The agent the consent created shows in Agents with the product’s name in the Harnesses column and OAuth in the Token column: it has no token to copy because the product holds the tokens. Connect a harness on its page gives the MCP URL and says to choose this agent under Connect as at the product’s consent, where a Hermes agent’s setup shows where its token goes.

​
Disconnecting

Revoke agent on the agent’s page ends the connection at once: the product’s next call is refused, its attempt to renew is refused, and it asks you to connect again. Connecting again is a new consent and, unless you pick the same agent, a new agent; the old one keeps its working set and history as your records. Disconnecting from the product’s side (removing the connector in Claude or ChatGPT) tells Graft nothing. The agent stays active with tokens nobody will present again; revoke it in the console when you are done with it.

​
Connecting the same product twice

Each connector you add is its own registration with Graft, and each consent is its own agent, so Graft in two Claude workspaces is two agents over one toolbox, exactly as two Hermes installs would be. If you want two products to share one agent, choose the existing agent on the second consent.

​
Prompts to try

​
Where next

Any MCP client

The two doors in protocol terms: the OAuth path this page uses and the static bearer path Hermes uses.

Agents

What the agent’s page shows for a connected product, and how to revoke it.