Connect ConvOps to OpenAI Codex

What do you get in Codex once ConvOps is connected?

Codex is good at long, self-directed coding runs. ConvOps gives those runs a process with checkpoints, so a person decides at the points that matter and every step is recorded.

1. Workflows Codex walks step by step. Each step arrives with its own instructions.
2. Gates evaluated by the engine. If a condition is unmet, Codex gets the reason back and the run stays put.
3. Approval steps that pause the run until you answer.
4. A memory shared with teammates on other clients, so a Codex session starts with what the team already learned.

ConvOps is the operations layer for AI agents: an MCP server that holds your team's process as workflows, with approval gates, a shared memory and an audit trail. It runs no AI models; Codex does the work. The steps, gates and decisions come from the ConvOps workflow engine, and the approvals and audit trail from governance.

What do you need to connect Codex to ConvOps?

A Codex setup that supports remote MCP servers, a ConvOps account, and the server URL. Plans start free; see pricing for limits.

How do I add the ConvOps MCP server to Codex?

Codex keeps MCP servers in ~/.codex/config.toml. Per OpenAI, the ChatGPT desktop app, Codex CLI and the IDE extension share that configuration, so you set it up once.

1. 1

   Add the server

   This writes a [mcp_servers.convops] entry to your Codex config. When Codex detects that the server supports OAuth, it starts the sign-in straight away and your browser opens the ConvOps sign-in.

   codex mcp add convops --url https://mcp.convops.app/

2. 2

   Sign in, if it did not start

   If the add step did not open a browser, or the token expired, start the OAuth sign-in yourself and approve access.

   codex mcp login convops

3. 3

   Or edit the config by hand

   The same entry, written directly. Useful when you manage dotfiles.

   [mcp_servers.convops]
   url = "https://mcp.convops.app/"

4. 4

   Check it in a session

   Start codex and type /mcp to see active servers. Then ask: "Start a ConvOps task for the dependency upgrade and follow its workflow."

   /mcp

export CONVOPS_API_KEY=ntk_...
codex mcp add convops --url https://mcp.convops.app/ \
  --bearer-token-env-var CONVOPS_API_KEY

Official reference: OpenAI: Codex MCP configuration, OpenAI: Codex and AGENTS.md. We checked these steps against those docs on Oct 4, 2026. OpenAI updates its interface often; when a menu name differs, the vendor docs are current.

What does a governed task look like in Codex?

A dependency upgrade. The first gate is a condition the engine checks, not a promise in the prompt: the advance is refused until the test step reports green. The merge step waits for a person.

Each tool row is Codex calling ConvOps. The engine hands over one step, checks the gate when the agent asks to advance, and holds the run at an approval step until a person answers. The same task can be picked up later, by you or by a teammate in another client, at the step where it stopped.

What should Codex users know about AGENTS.md and shared config?

Keep AGENTS.md about the code. Codex reads AGENTS.md files from ~/.codex and from the repository down to your working directory, and stops adding them at a combined size limit (32 KiB by default, per OpenAI). Process instructions that would push you toward that limit belong in a ConvOps workflow, where Codex receives only the current step. CLAUDE.md and AGENTS.md too long covers what to keep in the file.

One config, three surfaces. Because the CLI, the IDE extension and the ChatGPT desktop app share ~/.codex/config.toml, adding ConvOps once makes it available in all of them.

Unattended runs use other engines today. ConvOps executors dispatch unattended runs to Claude Code or OpenCode in an isolated pod, or hand them to your own runner. Codex is the attended client in that picture: you run the workflow in your session and answer the gates.

Codex can also be driven by other tools. Archon, for example, launches Codex per node of a YAML coding pipeline; the Archon comparison explains where each approach fits.

Why is ConvOps not showing up in Codex?

Almost always one of two things: the sign-in has not finished, or the server is off for this session. Start with the first row.

What else do people ask about ConvOps in Codex?

Short answers, checked against the OpenAI documentation linked above.