Coincide Get Coincide

Agent · OpenAI

Codex logo

Codex in Coincide

Coincide drives the Codex command-line tool you installed, through its app server, with your ChatGPT sign-in or OpenAI API key. You get Codex’s own sessions, plan mode and models, in a native Mac window next to your other agents.

How Coincide runs Codex

Coincide starts codex app-server and talks to it in JSON-RPC, the interface Codex provides for apps that embed it. Codex then talks to OpenAI (or to the provider it’s configured for) as it would in Terminal. Your ~/.codex configuration and AGENTS.md files keep working, because it’s the same tool. Coincide doesn’t proxy or meter anything.

Codex needs to be version 0.149 or later; Coincide asks you to update to 0.156 or later for full support. If codex isn’t on your shell’s PATH, set its location in Settings › Codex.

Set it up

  1. Install the Codex CLI (the official Codex CLI docs list other ways to install it):
    curl -fsSL https://chatgpt.com/codex/install.sh | sh
  2. Sign in with ChatGPT, or with an API key:
    codex login
  3. Check Settings › Codex in Coincide: it shows the version and account Codex reports, and can run the commands above in Terminal. Choose Check Again once you’re signed in.
  4. Choose a Codex model in the composer, or type /codex in a thread that another agent started.

Codex can sign in with ChatGPT, an OpenAI API key, Amazon Bedrock or a custom provider; Coincide shows which one is in use. To keep a second account, add it in Settings › Codex: a login account gets its own CODEX_HOME, and an API key is saved in your Keychain and given to that account’s Codex. See Agent accounts.

Models, effort and speed

Coincide reads the model list from Codex itself, so it shows exactly the models your account offers. For each one it offers the reasoning efforts that model supports, and the service tiers Codex reports for it (such as a faster tier) as speed options in the model picker. Image attachments are offered only for models that accept images.

What works with Codex

In CoincideWith Codex
Plan modeCodex’s own plan collaboration mode, with its plan updates shown in the thread.
Approvals and full accessAsk for approval runs Codex in a read-only sandbox and asks before anything that isn’t a known-safe read. Full access never asks and lifts the sandbox.
QuestionsCodex’s requests for input and MCP forms are answered inline. Link-style MCP requests are declined.
SteeringMessages you send while Codex works steer the running turn.
AttachmentsImages go to Codex as images; other files are mentioned by path.
SubagentsCollaborating agents show up in the turn, and their approval requests are labelled as a subagent’s.
ResumingEach thread resumes its own Codex thread, and unarchives it in Codex if needed. If Codex no longer has it, Coincide starts a new one and tells you.
StoppingStop interrupts the turn; if Codex doesn’t stop within five seconds, Coincide ends its process.
Usage limitsFor ChatGPT sign-ins, your plan’s rate limits and credits, updated live while Codex works. Optionally, credit details from chatgpt.com (off by default).
ImportingEarlier Codex sessions can be imported and continued.

Using Codex with other agents

A common pattern is to plan or review with one agent and implement with another. When you hand a thread to Codex, it gets a recap of the turns it missed and works in the same folder, and when you switch back the earlier agent resumes its own session. Switching agents explains what the recap contains.

Limitations