OpenAI Agents SDK
Give an agent built with the OpenAI Agents SDK, in Python or TypeScript, the Nextmail tools as an MCP server.
Install
pip install openai-agentsThe Python package needs Python 3.10 or later. Both read your OpenAI key from OPENAI_API_KEY, and an agent without a model uses the SDK’s default.
Sign in once
npx @nextmail/cli loginChoose the browser sign-in. On the approval page, pick the workspace, the access, the addresses the agent may reach and how long the approval lasts. Read only is the default, and it leaves out tools such as sendEmail, replyToEmail and createDraft, so choose Read and send when the agent should draft or reply. The sign-in is saved on this machine and listed in Account → Command line, where you can narrow it or remove it. A sign-in the bridge has not used for 7 days, or one whose approval has ended, has to sign in again.
On a server or over SSH, add --no-browser: open the link on any device, approve, and paste the code it shows back into the terminal. A browser sign-in needs a person once, so a CI job cannot make one.
The MCP server takes an approved sign-in and never an API key. A key sent to it gets 401, with a WWW-Authenticate header that points at the OAuth metadata. Code that holds a key calls the REST API or an SDK instead.
Run an agent
import asyncio from agents import Agent, Runnerfrom agents.mcp import MCPServerStdio, create_static_tool_filter async def main(): async with MCPServerStdio( name="Nextmail", params={"command": "npx", "args": ["-y", "@nextmail/cli", "mcp", "serve"]}, client_session_timeout_seconds=30, tool_filter=create_static_tool_filter( allowed_tool_names=["whoAmI", "listThreads", "getThread"] ), ) as server: agent = Agent( name="Inbox assistant", instructions="Read the mailbox with the Nextmail tools. Never send anything.", mcp_servers=[server], ) result = await Runner.run(agent, "Summarise my unread mail.") print(result.final_output) asyncio.run(main())The framework starts the local bridge, npx -y @nextmail/cli mcp serve, as a stdio MCP server. The bridge sends each call to the MCP server with your saved sign-in and renews the access token when it runs out, so the agent never handles a token. To use another profile of the command line, add --profile <name> to the arguments. The bridge runs on Node.js 20.12 or later.
- The static tool filter keeps the agent to the tools it names, by the names the server lists.
- In Python the client session timeout is 5 seconds by default, and the first
npxdownload can take longer, so the examples allow 30.
What the agent can do
The server registers up to 600 tools, and the agent sees the ones both your role and the approval allow. A tool the approval leaves out is never listed, so the model never tries it. Every tool loaded is sent to the model on each turn, so give the agent the ones its job needs. The example starts from whoAmI, listThreads and getThread.
| Tool | What it does | Access it needs |
|---|---|---|
| whoAmI | The role you hold in the active workspace, and every permission it grants | None |
| listThreads, getThread | List conversations with filters, and read the latest message of one | Read only |
| waitForEmail | Wait up to 60 seconds for a message that matches, such as a sign-in code | Read only |
| createDraft | Save a draft without sending it, as a reply when given a thread | Read and send |
| replyToEmail, sendEmail | Reply inside a thread, or send a new email | Read and send |
| modifyLabels | Add or remove labels on conversations | Read and send |
The MCP server never asks before a tool acts, so a tool that sends or deletes does it at once. A few tools make a change the REST API guards with a verification code, such as creating a rule or removing a domain, and the bridge cannot ask for a code. Run npx @nextmail/cli verify in a terminal with the same profile, or choose Allow changes for 60 minutes on this sign-in in Account → Command line, and those tools run for the next 60 minutes.
Without the command line
MCPServerStreamableHttp can connect to the remote server, https://api.nextmail.sh/mcp, with an OAuth provider from the MCP SDK: pass it as auth in Python (an httpx.Auth with MCP SDK 1, an httpx2.Auth with 2), or as authProvider in TypeScript.
The hosted MCP tool, which OpenAI runs on its side, takes an OAuth access token in authorization. Your app runs the sign-in, sends the token with every request, and renews it, since an access token lasts an hour.
Google Workspace