Human-in-the-loop for Codex
Codex can take a task and run with it, but a wrong assumption early costs you a whole PR review. Register Orako as a remote MCP server and Codex asks the person who actually owns the decision and keeps going, so what it ships matches what your team intended.
Connect Orako to Codex
- 01
Run
codex mcp add orako --url https://app.orako.io/mcp, or add the [mcp_servers.orako] block below to config.toml. - 02
Approve the connection with
codex mcp login orako. Codex loads theasktool at the same time. - 03
Human-only decisions go to the domain owner via Orako, and the run resumes on their answer.
You approve access inside Codex itself — no token to copy, nothing written to a local credentials file.
[mcp_servers.orako] url = "https://app.orako.io/mcp"
Project-scope config is honoured in trusted projects. Approve access with codex mcp login orako — OAuth happens inside Codex, no token to paste. The ask-first guidance ships with the server, not a local AGENTS.md block.
New to MCP in Codex? See OpenAI’s docs.
What Codex can ask
Give Codex a way to ask instead of assume.
Business rules before implementation
Codex is generating a proration routine and the rounding rule isn't written anywhere. It asks the billing owner and implements the confirmed rule, with the answer banked for next time.
Public API shape
Field naming on a public endpoint is a one-way door. Codex gets the API owner's call before the contract ships, not in review after the fact.
Intended behavior changes
A failing test could mean a bug or a changed requirement. Codex asks the feature owner which one it is before "fixing" the wrong side.
Codex + Orako, answered
Does Codex support MCP natively?
Yes. The Codex CLI loads MCP servers from config.toml under [mcp_servers.<id>] tables. Remote servers just need a url field, no local command or binary. User scope lives in ~/.codex/config.toml; trusted projects can carry their own .codex/config.toml.
How does Codex know when to ask instead of assuming?
Orako's MCP server sends its instructions to Codex the moment it connects: call search_history first, then ask only for decisions a human must make. There's no local AGENTS.md block to install or keep in sync.
Do my teammates need this agent installed to answer?
No. Only the developer running the agent connects Orako. The people answering receive each question in Slack, Discord, or the Orako dashboard and reply in the thread. They never touch an editor, a terminal, or an MCP config.
What happens to the answers?
Every resolved conversation becomes searchable team history with its authors, date, tags and full context. Agents call search_history before asking, so a question your team already answered can be reused instead of pinging anyone again.
Put a human in the loop for Codex
Give Codex the one thing it can’t generate: the answer only your team knows.