Connect an MCP server to your organization

Connecting your own client, such as Claude, Claude Code, ChatGPT or Codex, to Oddlynew is a different task: start with Compatibility and known limitations, then Setting up your client. This page is for organization admins who connect an outside MCP server so that every member's agent can use its tools.

Your organization can connect any remote MCP server once, and every member's agent can then use that server's tools through Oddlynew. Nobody installs anything in their own agent, and switching from one agent or model to another changes nothing: the connection belongs to the organization, not to a person or a chat client.

Connections of this kind are custom connections. Oddlynew has not reviewed the server or its tools and does not verify what they do. The page says so before you connect, and the connection is labelled that way afterwards. When a built-in integration covers an operation, agents use it and keep this connection for what it lacks, and a provider's money and bookkeeping writes stay on its built-in integration whenever it has one (Choose the route to an outside system).

What you need

You are an admin of the organization.
The server speaks MCP over Streamable HTTP at a public https address, on the current revision (2026-07-28) or on one of the 2025 revisions 2025-11-25, 2025-06-18 or 2025-03-26. On a 2025 revision every call starts a fresh session, so a server that remembers things between calls in one session forgets them. Servers on other protocol revisions, servers that run as a local process (stdio), and servers that need to ask questions mid-call are not supported; the check before connecting tells you when a server does not qualify.
One of three ways to authenticate: none (an open server), a fixed header such as an API key, or a sign-in with an account on the server (OAuth 2.1). Oddlynew keeps the sign-in and renews it.

Connecting, step by step

Open Settings, then Integrations, and choose MCP server.
Paste the server address and choose how it authenticates. For a header, enter the header name and its value and give the account a name that tells two keys for the same server apart. For a sign-in, you are sent to the server to sign in and come back.
Check the server. Oddlynew reads the server's name, protocol revision, instructions and tool list and shows every tool with the policy it will get.
Read the consent text. It answers four questions: who receives the credential (Oddlynew stores it encrypted and sends it only to this server, from Oddlynew's own servers; members and their agents never see it), what the connection can do (the server's tools, each with its policy), what Oddlynew verified (nothing; this is a custom connection), and who maintains it (your organization's admins, with the server's own documentation). It also states that every current and future member's agent acts as this account on the server, and that the connection stops working when you leave the organization.
Connect. The connection page opens with the tool list, the policies and the actions below.

Sign-in is selected by default; choose Header or None if your server needs a different method. You can also share a link that opens this form with the server address filled in, for example from your server's own setup guide: https://oddlynew.com/organization/connectors/mcp/connect?url=https%3A%2F%2Fmcp.higgsfield.ai%2Fmcp. url is the server address, percent-encoded. Two parameters are optional: auth=header or auth=none opens the form on that method instead of sign-in, and name shows a suggested name in the connection's name field, which the admin still fills in. Someone who is not signed in comes back to the filled form after signing in, in the organization they have open. The link only fills in the form: the address is checked as if typed, and nothing connects until an admin checks the server and confirms.

What a tool may do: read, write, blocked

Each tool has one policy, set by your organization. Read tools are called freely by Writer and Admin agents with read access; the Reader role cannot use connections. Write tools need write access and follow your organization's confirmation policy for their consequence class, the same way writes through Oddlynew's own integrations do; consequential ones wait for a person's confirmation by default (Integration approvals). Blocked tools are never called.

The starting policy comes from the labels the server gives its own tools: tools the server marks read-only become read; tools it marks as not destructive become writes of medium consequence; tools it marks destructive, or does not label at all, become writes of high consequence. Oddlynew cannot verify these labels; your decision to connect the server is the decision to trust them as a starting point. On the connection page you can tighten any tool to write or blocked, and your choice survives every later refresh. A tool the server adds later gets the same starting policy.

What members and their agents see

Every member's agent finds the connection through the MCP Servers Integration namespace in Oddlynew's run tool: list the connected servers, list or search a server's tools, read one tool's exact arguments, and call it. Every call is recorded in the organization's run log with the member whose agent made it, the connection and the tool. What a tool returns is data from a third party; agents treat it as information, never as instructions.

Keeping the connection healthy

Refresh catalog reads the tool list again. Do this when the server changed its tools; the page also tells you when a server refused a tool that Oddlynew still lists. When the server reports a tool as unknown, Oddlynew marks the catalog stale without blocking subsequent calls; refresh the catalog to reconcile it with the server's current tool list.
Reconnect appears when the server stopped accepting the credential or the sign-in; agents get a clear error until an admin reconnects. A sign-in is renewed by Oddlynew on its own as long as the server allows it.
Pause stops all calls without losing the setup; Disconnect removes the credential. Existing rules and data on the server are not undone.
Connecting the same server with a second account creates a second connection; agents tell them apart by connection, never by tool name.

Limits worth knowing

A connection holds at most 500 tools. One agent run may call a connection's tools at most 25 times, one call has a 25 second deadline, and one answer is bounded at 512 KiB. Calls count as provider calls under the ordinary run budgets; there is no separate fee for connecting a server.

Setting up your client explains how a person connects their own agent to Oddlynew, which is a different thing from connecting a server to an organization. Organizations and membership explains who is an admin and who is a member.