Connect an agent with MCP

Give your agent shared tasks and product context, with your own permissions and a record of its actions.

Endpoint: https://hub.n1.team/mcp · Streamable HTTP · stateless.

Connect with OAuth

In Claude, ChatGPT or another remote MCP client with OAuth support, add a custom server with the endpoint above. Availability depends on your client and plan.
Sign in to your existing Hub account, check the displayed workspace and approve access. Each person connects their own agent. You do not need to copy a key.
Send the verification prompt below. A saved connection alone does not prove the agent has reached the Hub.

For client implementers: discovery uses the 401 WWW-Authenticate resource_metadata link, protected resource metadata and authorization-server metadata. Authorization Code + PKCE S256, public clients, dynamic client registration, scope hub. No refresh-token grant is advertised: reconnect when authorization expires.

Connect with a personal key

For CLI clients and scripts, create a personal key in Agents → My connections. Store it as N1_HUB_KEY in the environment used by your client. The key is shown once; never commit it or place it in a URL. Revoke it from your own connections if exposed. Each key belongs to one person and workspace.

My keys

Codex: add to your MCP configuration, then launch Codex with N1_HUB_KEY available.

[mcp_servers.n1-hub]
url = "https://hub.n1.team/mcp"
bearer_token_env_var = "N1_HUB_KEY"

Other clients: choose HTTP / Streamable HTTP and configure the authorization header. Replace the placeholder using your client’s secret storage; environment substitution syntax varies by client.

Authorization: Bearer <your-personal-key>

Verify the connection

Call get_status. Tell me which person and workspace you act for. Then list_products and get_product_context for my product. Do not change anything yet.

The reply must identify your actual person and workspace. Check the first successful call in Agents. Product visibility matches your account; owner-only operations remain restricted. Removing a person or revoking a key stops subsequent requests.

Implementing a client: first protocol request

Use an MCP SDK to negotiate a supported protocol version, initialize, send notifications/initialized, then tools/list and tools/call. This diagnostic curl only initializes; it does not run a tool. Keep the negotiated version in the MCP-Protocol-Version header on subsequent requests. Stateless transport does not issue an Mcp-Session-Id and does not provide a persistent GET event stream.

curl --fail-with-body 'https://hub.n1.team/mcp' \
  -H "Authorization: Bearer $N1_HUB_KEY" \
  -H 'Content-Type: application/json' \
  -H 'Accept: application/json, text/event-stream' \
  --data '{"jsonrpc":"2.0","id":1,"method":"initialize","params":{"protocolVersion":"2025-03-26","capabilities":{},"clientInfo":{"name":"my-client","version":"1.0.0"}}}'

MCP client concepts · Streamable HTTP

Discover capabilities

The authenticated tools/list response is the complete live reference: names, descriptions, input JSON Schemas and annotations. Discover resources/list, resources/templates/list and prompts/list through the same client. Do not hardcode a tool count or assume every discovered action is allowed for your role.

Workflow	Start with
Identity and context	`get_status`, `list_products`, `get_product_context`
Tasks and handoff	`list_tasks`, `get_task`, `create_task`, `update_task`, `review_task`
Agent work queue	`next_task`, `get_agent_work`
Product data	`query_metric`, `get_overview`, `get_sources`
Updates	`get_digest`, `daily_digest` (prompt)

Product context is also available as n1://products/{slug}/context; plan_week is a planning prompt. Read the server instructions and agentPolicy before changing work. Closing a task requires stated grounds; taking work may require an estimate. A connection grants access; it does not start a background agent or run code. next_task must be called by your agent.

Limits and troubleshooting

401: missing, expired or revoked authorization. Follow OAuth discovery to reconnect, or check your personal key. Cookies are not MCP authorization.
429: up to 120 requests/minute per key, shared with REST API v1. Honor Retry-After; avoid tight polling.
413: request body exceeds 1 MiB. Reduce the payload.
Direct cross-origin browser requests are not supported; use a native or server-side client. A successful HTTP response can contain a tool result with isError:true. Inspect it before claiming success. Permission, validation and quota errors are returned by the tool; do not retry writes blindly.
GET /mcp returning 405 is expected for this stateless transport. Opening the endpoint in a browser is not a connection test. Use initialize and a real tool call.

For plain HTTP scripts, use the versioned REST API.