Connect your AI assistant to the Keyban MCP server

Tutorial

What you'll do

Wire an AI assistant — Claude, Claude Code, ChatGPT, Cursor or any other MCP client — to your Keyban organization through the Model Context Protocol. At the end you'll be able to ask things like "list my applications" or "create an agent wallet", and the assistant will execute them against your real organization, as you.

• Time~3 min
• DifficultyEasy
• You'll needA Keyban account

How it goes, end to end

1. 1

   Step 1: Add the server

   Give your assistant the Keyban MCP server URL.

3. 2

   Step 2: Sign in

   Authorize the assistant with your Keyban account (OAuth).

5. 3

   Step 3: Verify

   Ask the assistant to list your applications.

Three phases: add the server URL to your assistant, sign in with your Keyban account, then verify with a real prompt.

The server

The Keyban MCP server is hosted. There is nothing to install and no key to copy: the only thing your assistant needs is the URL.

https://mcp.keyban.io/mcp

The server authenticates with OAuth 2.1. The first time your assistant connects, it opens a Keyban sign-in page in your browser; you sign in with the account you use on the Keyban dashboard and authorize the assistant. Every tool call then runs as you, with your permissions in your organization, and appears under your name in the audit log.

Running the platform locally?

With tilt up -- --mcp, the same server answers on https://mcp.keyban.localtest.me/mcp and signs you in against your local admin. Replace the URL in every snippet below.

Step 1 — Add the server to your assistant

Pick the section matching your assistant. Each one ends with the same result: a keyban server registered, waiting for you to sign in.

Claude (claude.ai, Claude Desktop, Cowork, Claude Code)

Add to Claude

1. The link opens Customize → Connectors on claude.ai with the Add custom connector dialog already filled in: name Keyban, URL https://mcp.keyban.io/mcp. Without the link, open that page, choose Add custom connector and enter them yourself.
2. Click Continue, keep the detected options (Sign in now and Register automatically), then click Add.
3. Click Connect on the new connector, sign in with your Keyban account and authorize the organization.

The connector belongs to your Claude account, so Claude Desktop, Cowork and Claude Code pick it up as well: in Claude Code, /mcp lists it as claude.ai Keyban.

On Team and Enterprise plans an owner adds the connector once from Organization settings → Connectors, where this link prefills the same dialog; members then connect their own account from Customize → Connectors.

Claude Code

Signed in to Claude Code with your claude.ai account? Add the connector in Claude and you are done. Claude Code only skips claude.ai connectors when it runs with an Anthropic API key, Amazon Bedrock or Google Vertex; in that case, register the server yourself:

claude mcp add --transport http --scope user keyban https://mcp.keyban.io/mcp

Then, inside a Claude Code session, run /mcp, pick keyban and follow the sign-in in your browser. claude mcp list shows the server from the shell; claude mcp remove keyban removes it.

ChatGPT

1. Open Settings → Security and login and turn on Developer mode. Its availability depends on your account and, in a workspace, on its policy.
2. Open the app creation form, or go to ChatGPT Plugins and select Create app, then Create MCP app. Without developer mode, the page only offers to create or import a plugin.
3. Name it Keyban, add a short description, paste https://mcp.keyban.io/mcp under Connection and click Create.
4. Sign in with your Keyban account and authorize the organization.
5. In a new chat, open + → More and select the Keyban plugin.

Codex, including its side of the ChatGPT desktop app, uses its own MCP configuration: set it up as described in the next section.

Codex

codex mcp add keyban --url https://mcp.keyban.io/mcp

Then run codex mcp login keyban and sign in from your browser. The ChatGPT desktop app, the Codex CLI and the Codex IDE extension share this configuration (~/.codex/config.toml). From the desktop app, Settings → MCP servers → Add server with Streamable HTTP and the same URL does the same thing; select Authenticate afterwards.

Cursor

Install in Cursor

The button opens Cursor and registers the server. To do it by hand, add this to .cursor/mcp.json (project) or ~/.cursor/mcp.json (global):

{
  "mcpServers": {
    "keyban": {
      "url": "https://mcp.keyban.io/mcp"
    }
  }
}

Cursor then shows a Needs login prompt on the server; click it to sign in.

VS Code

Install in VS Code

Or from the Command Palette: MCP: Add Server → HTTP, URL https://mcp.keyban.io/mcp, name keyban. Start the server from MCP: List Servers and allow the authentication when VS Code asks.

Other clients

Client	Where the URL goes
Windsurf	mcp_config.json, entry "keyban": { "serverUrl": "https://mcp.keyban.io/mcp" }
Zed	settings.json, under context_servers, as an HTTP server with that URL
Gemini CLI, Gemini Code Assist	~/.gemini/settings.json; if your version has no native OAuth for HTTP servers, wrap the URL with npx mcp-remote https://mcp.keyban.io/mcp
Goose	Add extension → Remote (Streamable HTTP) with the URL
Raycast	Install Server command, transport HTTP, the URL

For anything else, use https://mcp.keyban.io/mcp as a Streamable HTTP server and let the client run its OAuth flow. Most clients accept this configuration shape:

{
  "mcpServers": {
    "keyban": {
      "type": "http",
      "url": "https://mcp.keyban.io/mcp"
    }
  }
}

To register the server in every coding agent installed on your machine at once, the third-party add-mcp tool does it in one command:

npx add-mcp https://mcp.keyban.io/mcp

Clients without OAuth (headless runtimes, CI)

A client that cannot open a browser can still authenticate with a Keyban API key: mint one from Settings → API keys on the dashboard and send it in an X-Api-Key header on every request. The key acts for the whole organization with the permissions it was given, not as a specific user, so prefer OAuth wherever the client supports it.

claude mcp add --transport http --scope user keyban https://mcp.keyban.io/mcp \
  --header "X-Api-Key: $KEYBAN_API_KEY"

Step 2 — Sign in and choose the organization

The sign-in page is the Keyban admin. After you authenticate, a consent screen names the client asking for access and lists what it will be able to do.

If you belong to a single organization, it is selected for you. If you belong to several, the screen asks which organization the assistant may act for: the access token carries that choice, and every tool call runs in that organization only. To switch, remove the server from your assistant and add it again.

Step 3 — Verify with three short prompts

Open a new chat (or run claude for Claude Code) and try these prompts in order. Each one exercises a different layer.

3.1 — Read the organization

List the applications in my Keyban organization.

The assistant calls list_applications. A fresh organization returns an empty list, which is a valid answer. A 401 means the sign-in did not complete — see Troubleshooting.

3.2 — Run a tool with elicitation

Look up the application called "Loyalty Portal" in Keyban.

The assistant calls lookup_application_id. If several applications match, the server elicits a choice — the assistant shows you a small picker — and resumes with your answer.

3.3 — Optional: a write call

Create an application called "Demo" in my Keyban organization.

The assistant invokes create_application. The new application appears on the dashboard within seconds. This proves the server handles tool calls (POST endpoints), not only resources. It is a real application in your organization: delete it from the dashboard afterwards if you do not need it.

LLM endpoints are not exposed via MCP

You won't find llm_chat_completion or list_llm_providers in your assistant's tool palette. Your assistant is already an LLM; routing chat completions through Keyban's LLM API would be redundant and double-billed.

Troubleshooting

The sign-in page never opens. The client did not detect the OAuth configuration. Check the URL ends with /mcp, remove the server and add it again. From a shell, curl -i https://mcp.keyban.io/mcp must answer 401 with a WWW-Authenticate header pointing at /.well-known/oauth-protected-resource.

Every tool call returns 401. The authorization expired or was revoked. In Claude Code run /mcp and sign in again; in Claude open the connector and click Connect; in ChatGPT open the plugin from ChatGPT Plugins; in Codex run codex mcp login keyban.

Org-scoped tools fail although you are signed in. You belong to several organizations and none was chosen at consent time. Remove the server, add it again and pick the organization on the consent screen.

The connector is saved but the tools do not appear. Start a new chat: Claude and ChatGPT refresh the tool list per conversation.

Claude Code: /mcp shows the server as failed. Run claude mcp list to inspect the registered configuration, fix the URL with claude mcp remove keyban then claude mcp add again.

Going further

• Ask the assistant to read the Keyban guide: it calls get_instructions and follows the keyban-admin skill, which takes a new account all the way to an installed agent wallet MCP.
• The MCP tools reference lists every tool, resource and prompt the assistant can now call, including what is intentionally excluded.
• The MCP Inspector validates the connection and the OAuth flow without involving an LLM client.