Use okdata in Codex: social data through MCP

With this set up, Codex can check real data before it writes code that depends on it. Picture a weekly report on how viewers react to a competitor's YouTube launch video. You ask Codex to fetch one page of comments through okdata, look at what comes back, then write a small script that does the same every Monday over the plain HTTP API and saves a CSV. The exploring happens over MCP; the script it leaves behind needs no AI at all.

Before you start

Codex installed: the CLI (codex runs in a terminal), the IDE extension, or the ChatGPT desktop app. All three read the same configuration, so you set this up once.
An API key from okdata, starting with okd_. New keys come with $0.10 of credit. The quickstart explains how to get one.
The key in the OKDATA_API_KEY environment variable. On a Mac, add this to ~/.zshrc (Linux: ~/.bashrc) and open a new terminal:

export OKDATA_API_KEY="okd_..."

Add the server

Open ~/.codex/config.toml (create it if it doesn't exist) and add:

[mcp_servers.okdata]
url = "https://api.okdata.online/mcp"
bearer_token_env_var = "OKDATA_API_KEY"

bearer_token_env_var names the variable, not the key. Codex reads OKDATA_API_KEY when it connects and sends Authorization: Bearer <key>, so the key never lands in the file. To limit the server to one project, put the same lines in .codex/config.toml inside that project instead; Codex only reads project files in projects you have marked as trusted.

In the IDE extension and the desktop app you can also add servers from the settings (MCP servers, then Add server, choosing Streamable HTTP), but the bearer-token setting lives in config.toml, so editing the file is the quickest way. Restart the extension or the app afterwards.

Check that it worked

From a terminal, codex mcp list should show okdata with its URL. Inside a Codex session, type /mcp to see it connected with four tools: find, describe, call and account.

Then ask: "What is my okdata balance?" A number back means the key works, and the check is free.

A first prompt to try

Use okdata to find the YouTube video comments endpoint and tell me its price.
Fetch one page of top comments for https://www.youtube.com/watch?v=<video id>.
Then write comments_weekly.py that makes the same request over plain HTTP,
reads the key from the OKDATA_API_KEY environment variable, and saves a CSV.

What it costs

find, describe and account are free. Each call costs the endpoint's price. One page of YouTube comments is $0.004, so the prompt above costs $0.004 plus whatever you later spend running the script. A page of Xiaohongshu note search is $0.02. Failed calls are free.

If you run Codex with broad auto-approval, include a budget in the prompt, such as "no more than two paid calls".

When something goes wrong

unauthorized (401). Codex could not read the key. Start Codex from a terminal where [ -n "$OKDATA_API_KEY" ] && echo set || echo missing prints set. Apps opened from the Dock or Start menu may not see variables set only in your shell profile, so the desktop app can fail where the CLI works. As a last resort, replace the bearer_token_env_var line with http_headers = { "Authorization" = "Bearer okd_..." }; the key then sits in plain text in ~/.codex/config.toml, so never do this in a project's .codex/config.toml.
insufficient_balance (402). Not enough balance for the call; nothing was charged. Top up and retry. Prices are on the pricing page.
rate_limited (429). Too many calls at once. The error includes retry_after_seconds; wait and repeat the same call. No charge.
The server or its tools are missing. Check the TOML: the section header must be [mcp_servers.okdata] with an underscore, and the URL must end in /mcp. If the server is slow to start on a poor connection, add startup_timeout_sec = 30 under it.