Two ways to use Codex for coding:
Upload files, run code in a sandbox, schedule tasks. Quick and self-contained.
Sees your whole workspace: AGENTS.md, context files, source lists. Builds where your context lives.
Your assistant
isn't limited to
what's in your
folder.
MCP (Model Context Protocol) = a standard for connecting AI assistants to external tools. Think of it as apps for your assistant.
| MCP Server | What it does | Newsroom use case |
|---|---|---|
| Chrome DevTools | Navigate, click, screenshot, inspect DOM & network | Capture & inspect how a page loads, debug a prototype |
| Playwright | Automate browsers via accessibility tree | Collect data from public websites on a schedule |
| YouTube Transcript | Pull transcripts from any YouTube video | Get full text of a press conference or hearing |
| NotebookLM | Query your notebooks, get citation-backed answers | Research a document trove with grounded citations |
| Read, search, send messages & media | Send story alerts to a group, search past conversations | |
| Google Drive | Search, read, create files; Docs → markdown, Sheets → CSV | Pull a shared doc, read a spreadsheet, update a brief |
| UNHCR ← today | Query refugee stats by country, year, demographics | Get displacement data for a story without manual download |
Connect your assistant to live UN refugee data. One install. Then ask it real questions about displacement worldwide.
github.com/rvibek/mcp_unhcrYou described the goal
AI called UNHCR API
Saved to your workspace
Codex can now spawn specialized agents that work in parallel and consolidate results, without you directing each step.
One orchestrator.
Many workers.
Fresh context each.
Three built-in agents ship with Codex:
Read-heavy analysis: maps documents, traces code paths, researches background
Execution-focused: implements tasks, processes batches, writes outputs
General-purpose fallback for anything not assigned to a specialist
You can also build your own →
One TOML file.
One specialist.
Drop a .toml file in .codex/agents/
"Investigate this company. Have explorer map all public filings, worker extract financials, and default write the brief."
Give it a list of 5 sources. One worker per source, all running in parallel. Results consolidated into one report.
"Check this draft. Have explorer verify every factual claim and worker find conflicting sources. Flag anything uncertain."
One agent per publication, running in parallel. "How are these 10 outlets covering this story differently?"
Up to 6 parallel threads by default · Each agent gets a fresh context window
| MCP | Skills | Subagents | |
|---|---|---|---|
| Kitchen | Ingredients | Recipe | Sous-chefs |
| What | External data & tools | Reusable instructions | Parallel autonomous agents |
| Difficulty | Medium (install) | Easy (markdown) | Medium (prompt pattern) |
| Use case | Search, archives, CMS | Interview prep, digest | Batch research, investigation |
Every AI output needs editorial review before it goes live
If it can't show you where it found something, be skeptical
Ask it to say "I'm not sure" instead of filling gaps with plausible-sounding text
Source identities, embargoed stories, off-the-record material — never in AI tools
Use your assistant daily for 30 days.
Add one thing per week — a file, a skill, a tool.
Note what works. Share wins with your team.
AGENTS.md + context files + skills that know your beat
A reusable workflow you'll use before every interview
Codex-built Python tool — without writing Python
Your assistant now talks to the world outside your folder
Parallel research with named specialists — just shipped in Codex