3.2 MCP: Plugging In GitHub, Figma & More

At some point you'll catch yourself doing this: pasting an error from Sentry into a Codex session so it can debug it, then pasting a different one from the same dashboard three days later because the context from last time is gone. Or explaining your team's ticket format for the fourth time. That repetition is a signal, and MCP is the fix.

MCP — Model Context Protocol — is the standard that lets Codex talk directly to things outside your codebase: GitHub, Figma, Sentry, a database, an internal company API. Connect an MCP server once, configured in config.toml, and Codex can query it as a normal part of its workflow from then on, the same way it reads a file — no more copy-paste round trips, no more re-explaining context that actually lives somewhere else. Recent versions added per-tool output token limits specifically so a chatty MCP server doesn't quietly eat your whole context budget on one call, and credential handling for anything requiring OAuth is encrypted locally rather than sitting around in plain text.

The practical test for whether something should be an MCP connection instead of a copy-paste habit: are you explaining the same external context more than once? Your ticket system's format, your error-tracking dashboard, your design system's token names in Figma. The first time is a conversation. The third time is a connection you should have set up after the first.

For a team rather than a solo user, the same logic applies at a larger scale — an organization can standardize which MCP servers everyone connects to, instead of each engineer configuring the same integrations by hand and inevitably drifting out of sync on versions or auth.