Original video in Chinese.
Key Takeaways
- X MCP gives Codex a source of live information through the official X API. It can retrieve account timelines, search posts, track users, and summarize themes, changing views, and recurring keywords.
- My AI OS has three layers: accumulated knowledge, an agent workspace, and live external signals. X MCP connects the third layer to the first two through a structured, monitorable data pipeline.
- Install the MCP, complete OAuth authorization, and start with a small task, such as collecting discussions of loop engineering into a topic brief. Its lasting value is a reusable intelligence system for important accounts and trends.
Today I finally connected X to Codex through the official MCP, so that Codex can call the X API and retrieve data directly.
When I find an important account, I can ask Codex to retrieve its posts from the last month and summarize the main areas of interest, shifts in opinion, and recurring themes. That is more useful than reading isolated posts because I can see the account’s overall focus during that period.
This tool matters to me because my AI OS has three layers. The first is a knowledge base for long-term accumulation. The second is a workspace where agents perform tasks. The third is a source of real-time signals that tells agents what is happening outside.
The first two are largely in place. The missing part is the third, and X is the most important source for it. Almost all the AI and US stock-market information I follow comes from X.
The problem is the volume. Useful and useless information, true and false claims, all arrive together at high speed. Even after spending considerable time, I can still miss what matters.
Many people, myself included, want AI to help search and organize that information.
X already offered an API, but integrating it required handling OAuth login and token refresh, wrapping the API requests, and exposing those capabilities to AI.
The official X MCP packages OAuth, token refresh, tool definitions, and agent integration. For the first time, my AI OS has a stable input for live signals: X is the radar, Codex is the workspace, and MCP is the pipeline connecting them.
That makes it possible to observe the outside world, organize signals, and form judgments within a working system.
The four setup steps I followed
First, give Codex—or another agent—the official MCP documentation and ask it to install the MCP and X API integration. Let the agent handle that part. If it cannot complete it, try another agent.
Second, visit the X developer dashboard, create an account, add a payment method, and fund it. I started with $5.
Reading X data through the API costs money. In my test, reading my own data was relatively inexpensive, at roughly $1 per 1,000 resources. Public posts, users, and search results had different resource prices.
This is best used for valuable monitoring, such as following important accounts, searching keywords, and organizing trends, rather than indiscriminate bulk collection.
Third, create an application in the developer dashboard. Following Codex’s instructions, enter the callback URL in the settings and provide the Client ID and Client Secret for the integration.
Finally, open the authorization link, grant authorization, and return the resulting full URL to Codex. That completed the setup in my test.
Through the official MCP, Codex could read timelines, search posts, look up users by username, and manage bookmarks.
At the time of this test, the MCP did not support posting. I tested posting through xurl CLI, which calls the X API directly, and it worked. For longer content, it could split the text into a thread.
For an occasional lookup, Grok may be enough. The real value of X MCP comes from making X a data source inside your AI workflow and a component of your AI OS’s perception layer. That is when spending money on the X API becomes worthwhile.
Start with one small task
Do not begin with a complex automation. Try this:
Use X MCP to search discussions of loop engineering on X. Organize the main topics, disputes, and representative people into a topic brief.
If that works, you already have a small X intelligence system.
That is all for this episode. If you want to build your own AI OS and bring AI into your information flow, knowledge base, and workspace, join the newtype community. See you next time.