How to Use Codex in Slack with Your Whole Team

Dott

Someone on your team just asked if you can put Codex in Slack so non-engineers can ask questions about the codebase and engineers can kick off small fixes from a thread. If your engineers already live in OpenAI's coding agent (or split their time across Codex and Claude Code), this sounds like a five-minute app install. In practice, you end up choosing between three different architectures: OpenAI's official Slack app, a homemade bot on a server, or bridging Slack to the Codex CLI already running on your engineers' machines.

The deciding question for a team is whose Codex account runs each request. That one detail controls who in the company is allowed to ask a question, whether a teammate can reply in the same thread to steer the work, what repositories and local tools the agent can reach, and how many seats or API tokens you pay for at the end of the month.

By the end, you'll have a working Slack setup where your team can mention an agent, run real repository work through Codex, and get answers or pull requests back in the thread. Expect setup to take about 30 minutes for the official app or a local machine bridge, and an afternoon if you wire up your own bot.

Key takeaways

  • The official bot is now @ChatGPT: OpenAI moved Slack requests from @Codex to @ChatGPT, which delegates coding questions to a separate Codex Cloud task.
  • Official app requests run on the requester's own account: Every person who mentions @ChatGPT in Slack needs their own ChatGPT account with Codex Cloud access, a linked Slack connection, and access to a published cloud environment.
  • Sharing one personal login on a server breaks OpenAI's terms: If you build or self-host a bot for a channel, authenticate it with an OpenAI API key (and pay per token) or Enterprise access tokens.
  • Running Codex on teammates' machines keeps local tools and plans: Tools like Merlin route Slack mentions to Codex on your engineers' computers, keeping usage on the machine owner's plan and giving the agent access to local repos and MCP servers, provided the machine is awake.
  • Pick your setup by who asks and who pays: Use the official app if everyone already has a ChatGPT Business seat, an API-key bot for a single channel experiment, or teammates' machines if you want the whole company asking questions against real dev environments.

What Codex in Slack does today

If you search for a Codex Slack bot, most tutorials still tell you to mention @Codex. OpenAI has since consolidated its Slack app under @ChatGPT (Use ChatGPT in Slack, seen on November 11, 2026). When you mention @ChatGPT in a channel and Cloud delegation is enabled, ChatGPT hands repository work off to a separate Codex Cloud task and posts the result back to the Slack thread.

OpenAI's Use ChatGPT in Slack docs page: mention @ChatGPT in a Slack channel, set up the Slack app as a workspace admin, Cloud delegation requires a published environment shared with the workspace, and connect your account when promptedThe official Codex Slack integration now answers to @ChatGPT and sends repository work to a separate Codex Cloud task.

That cloud handoff shapes everything about how a team uses it. Each coding request spins up a fresh virtual machine in OpenAI's cloud (2 vCPU and 8 GiB of RAM on Plus, or 4 vCPU and 16 GiB on Pro, Business, and Enterprise, per the cloud environments docs). Because it runs in a fresh cloud box, it continues working even if you close your laptop. It also means your local skills, local MCP servers, and self-hosted GitHub Enterprise Server aren't there.

The biggest catch for a team is how accounts work. In OpenAI's app, the coding task runs as the requesting user. Every person who wants to ask @ChatGPT a question in Slack needs their own ChatGPT seat and has to link their Slack account. When a second teammate replies in the same Slack thread to add context, OpenAI's docs note that a request using another account may start a separate coding task. You hit a similar wall with Claude Code in Slack when trying to turn a single-player cloud sandbox into a multiplayer team thread.

Three ways to set it up

Before you touch any admin settings, compare the three ways to wire Codex into Slack. Look at the second and sixth rows first: whose account runs the request, and what happens when a teammate replies in the thread.

Official @ChatGPT appHomemade botCodex on teammates' machines (for example Merlin)
Where Codex runsCodex Cloud, one task per requestOne host you run (a server or spare Mac)The machines of the engineers who connect them
Whose account runs a requestThe requester's own ChatGPT accountOne API key (a shared personal login breaks OpenAI's terms)The machine owner's own Codex login
What you payA ChatGPT seat per person who asks (Business $20 to $25 per user per month) and their usageAPI tokens per task, plus hosting and your timeEach engineer's existing Codex plan, plus Merlin from $9 per workspace per month
Who can askPeople with a seat and a linked accountAnyone in the channelWorkspace members, on runtimes their owners opened to them
What it can reachRepos in a published shared cloud environment and its secretsWhatever is on the hostThe owner's local repos, tools, MCP servers and credentials
A teammate replies in the threadStarts a separate task under their accountWhatever you buildSteers the same task
LinearSeparate @Codex integrationBuild it yourselfSame agent, delegated from Linear
MemoryNone at launch, AGENTS.md onlyBuild it yourselfPer-channel team memory, opt in
Main catchTwo admins, a seat and a linked account for every requesterYou own the Slack plumbing, retries and securityMachines must be online, and two bills

What you need

Depending on which path you pick below, gather these pieces before you start:

  • For the official @ChatGPT app: A ChatGPT workspace on a Business ($20 per user per month billed annually, or $25 monthly) or Enterprise plan (Codex pricing, seen on November 11, 2026), a ChatGPT workspace admin, a Slack workspace owner or app manager, and a GitHub cloud repository configured as a shared Codex Cloud environment.
  • For a homemade bot: A server or always-on machine with Node.js or Python, the Codex CLI installed (codex --version), a Slack app permission to install bots in your workspace, and an OpenAI API key with billing enabled.
  • For Codex on teammates' machines: At least one engineer machine (macOS, Windows, or Linux) with the Codex CLI installed and signed in (codex login), a Slack account able to install apps, and a Merlin workspace (14-day free trial, then from $9 per month).

How to set up @ChatGPT

Use this path if every person who will talk to the bot in Slack already has a paid ChatGPT seat on your company workspace and your repositories live on cloud GitHub.

1. Check the plan and admins

Check that your team is on a plan that includes the Slack cloud integration and that your requesters have enough usage headroom. While OpenAI's pricing table lists the Slack integration on Plus, Pro, Business, and Enterprise, the admin deployment docs walk through a workspace-level rollout managed from the OpenAI admin console, which means Business or Enterprise in practice. Keep in mind that local CLI messages, ChatGPT Work, and cloud tasks all draw from the same usage pool, and cloud tasks burn through allowances faster than local messages.

Codex pricing page with the notice that ChatGPT Work and Codex share usage, and Free and Go plan cards at $0 and $8 per monthSlack requests draw on the same allowance as everything else the requester does in ChatGPT and Codex.

Make sure you have both a ChatGPT workspace admin and a Slack workspace owner (or app manager) available, because one starts the connection and the other approves it.

2. Connect Slack

Have your ChatGPT workspace admin open the OpenAI admin console, select your workspace, open Agents, and choose to connect Slack. This creates one surface per Slack workspace (if you run Slack Enterprise Grid, each grid workspace gets its own surface). Next, have your Slack admin approve the requested permissions and install the app. If your workspace previously installed the older @Codex app, your Slack admin still has to approve the updated permissions before @ChatGPT will respond.

Watch out when restricting channels: choosing Selected channels only limits where the bot works in public and private channels, and does not block direct messages.

3. Share a cloud environment

Installing the Slack app only gives you chat, and coding work takes a separate switch. In the admin console, enable Codex Cloud for Slack. Then, set up a cloud environment for your repository and publish it to the team: open the environment settings, go to Privacy, and change Who can use from Only me to your workspace. Personal cloud environments are ignored by the Slack integration.

If you turn on role-based access control for Codex Cloud, grant access to both the Slack surface's service account (which finds environments) and every individual user who will make requests.

Ask each teammate who wants to use the bot to mention @ChatGPT in an enabled channel. The first time they do, @ChatGPT replies with a connection link. They click the link, sign in to their ChatGPT account for that workspace, and then mention @ChatGPT a second time in the original Slack thread to kick off the actual work.

5. Name the repository

Start a thread by naming the exact repository or shared environment so the service account picks the right sandbox:

@ChatGPT investigate the failing checkout tests in acme/storefront and propose a fix.

To follow up on that work from the same user account, reply in the same thread:

@ChatGPT add a regression test for the fix and summarize the checks you ran.

A private Allow or Deny prompt may appear for the requester, along with Follow along and Cancel buttons. Once the task finishes, review the changes in Codex Cloud before committing or opening a pull request. Because @ChatGPT in Slack has no persistent memory across threads, put any standing instructions for the repo inside an AGENTS.md file in the repository root.

How to build your own bot

If you have non-engineers without ChatGPT seats who need to ask questions in one channel, or you need Codex to reach a private network that Codex Cloud can't see, you can run a bot on your own server or spare Mac.

1. Choose an allowed login

Before you write any code or clone a GitHub bridge, decide how the bot will sign in to Codex. Many weekend tutorials tell you to run codex login with your personal ChatGPT Plus or Pro account on a Linux box and let the whole company share it through Slack. Both OpenAI's Terms of Use and Business Terms forbid sharing individual login credentials across multiple users.

For a shared bot, OpenAI's authentication docs state that API keys are the recommended default for automation (Enterprise workspaces can also mint Codex access tokens for scripts and runners). Export an API key on the host machine. On API pricing (seen on November 11, 2026), gpt-6.1-sol costs $2.00 per million input tokens ($0.10 cached) and $10.00 per million output tokens, while gpt-6-astra runs $10.00 input ($1.00 cached) and $50.00 output. As a rough illustration, a task that reads 500K uncached input tokens and writes 30K output tokens on gpt-6.1-sol costs about $1.30, with prompt caching dropping repeat reads significantly.

2. Create the Slack app

In the Slack API dashboard, create a new app for your workspace and turn on Socket Mode. Socket Mode connects outbound over a WebSocket using an app-level token, so you don't have to expose a public HTTP endpoint on your server. Under OAuth & Permissions, add the app_mentions:read and chat:write bot token scopes, subscribe to the app_mention event under Event Subscriptions, and install the app to your workspace.

3. Acknowledge, then run Codex

Slack's Events API expects an HTTP 2xx acknowledgement within three seconds, or it assumes your bot died and retries the event up to three times. In your event handler, acknowledge the mention immediately, post a "working on it" reply, and spawn the Codex run asynchronously.

You have two ways to drive local Codex from code (skip the Codex app server, whose docs note that its WebSocket transport is experimental and its auth isn't permitted for hosted services):

  • The CLI in non-interactive mode: Run codex exec with --json to stream JSONL progress events and --sandbox workspace-write if you want it to edit files (the default sandbox is read-only, and the old --full-auto flag is gone).
  • The official SDK: Install the Codex SDK (@openai/codex-sdk in TypeScript or openai-codex in Python) and call codex.startThread() and thread.run().
codex exec --json --sandbox workspace-write "investigate the failing checkout tests and propose a fix"

4. Map threads to sessions

So teammates can reply in a Slack thread without starting from scratch, store a mapping in a small SQLite table or JSON file between Slack's thread_ts timestamp and the Codex session or thread ID. When a new message arrives in an existing thread, resume that session instead of starting a new one:

codex exec resume "$SESSION_ID" --json --sandbox workspace-write "add a regression test for the fix"

In the TypeScript SDK, pass the stored ID to codex.resumeThread(threadId).

If you don't want to write the Slack Socket Mode and session-mapping code yourself, open-source bridges like cc-connect (16.3k stars on GitHub as of November 11, 2026) or smaller repos like earonesty/codex-slack already wire Slack threads to local Codex sessions over Socket Mode.

GitHub repository page for cc-connect with 16.3k stars, described as a bridge from local AI coding agents including Codex to messaging platforms such as SlackOpen-source bridges like cc-connect run Codex on one host machine, under whatever login that machine has.

Just watch how you authenticate that host: every open-source bridge runs all Slack users through the single host's Codex environment, so configure that host with an API key.

How to use teammates' machines

If your engineers already pay for ChatGPT Plus, Pro, or Business plans and have Codex set up locally with their repositories, credentials, and MCP servers, you can connect those machines directly to Slack with Merlin. Instead of spinning up a blank cloud VM or maintaining a shared bot server, requests route to an engineer's machine and run on that machine owner's Codex login.

1. Install Merlin

First, check that Codex works in your terminal (prerequisites):

codex --version
codex login

Download the Merlin desktop app (macOS or Windows) or the Linux CLI daemon and sign in to your Merlin workspace. Merlin detects the local Codex installation and registers your computer as a device without asking for an OpenAI API key.

2. Set runtime access

Open the Merlin web app and go to Settings > Devices (devices and runtimes) to see your connected computers, check their available runtimes (Codex and Claude Code), and pick your preferred runtime. By default, a new runtime is set to private, meaning only you can trigger it from Slack.

Web app Settings Devices page listing a connected MacBook Pro with Claude Code and Codex available, a Preferred runtime selector set to Claude Code on that MacBook, and Install Merlin stepsIn Merlin each machine reports its Codex and Claude Code runtimes, and each person picks the one their requests should prefer.

If you want non-engineers or teammates to run requests on your machine, open Agent in the web app, pick your device, and open the Codex runtime page. From there, flip its access setting to public, pick the default model and reasoning effort, and set concurrency (from 1 to 16 parallel tasks, with 4 as the default). Keep in mind that when a teammate runs a task on your public runtime, it uses your local files and tools and draws from your Codex plan's usage allowance.

3. Connect Slack channels

In the Merlin web app, connect your Slack workspace (Slack guide). Add public channels from the Slack integration page, and invite the bot to private channels by typing /invite @Merlin in the channel. Each channel has three response modes: Mention only, Thread follow-ups (unmentioned replies in an existing Merlin thread may continue work), and Proactive (it may chime in when it can help). Thread follow-ups and Proactive only take effect while a workspace owner or admin has the workspace Proactive responses setting turned on.

4. Ask in a thread

Any workspace member who can use an available runtime (their own, or one its owner made public) can now mention @Merlin in a channel to ask a question or ask for a code change. You can also ask for a specific runtime or effort level right in the message (for example, "use Alice's" or "use max effort").

Slack thread in the engineering channel: Mia asks @Merlin to find and fix a trial billing bug, Merlin's task card shows the task complete, Sam adds that annual plans need the same fix, and Merlin's final answer covers bothA teammate's follow-up in the thread steers the same task instead of starting a new one.

Unlike the official @ChatGPT app, when a second teammate replies in the same Slack thread with extra context (like Sam pointing out that annual plans need the same fix in the screenshot above), their reply steers the running task instead of forking a separate session under a second account.

5. Add memory and Linear

If you want the agent to remember facts across threads without editing AGENTS.md every time, an owner or admin can enable memory, which saves short, durable notes scoped to each Slack channel (public channel notes help the whole workspace, while private channel notes stay in that channel). You can also connect Linear to the same Merlin workspace so delegating a Linear issue runs on the exact same Codex machines, and Linear tasks read the same public-channel memory as Slack tasks when memory is on. By contrast, OpenAI's Linear integration is a separate @Codex app running on older cloud environments that doesn't share setup or memory with @ChatGPT in Slack.

The trade-offs to plan around with Merlin: at least one connected machine has to be awake and online to take work, your team has two bills (each engineer's own Codex plan plus Merlin's workspace plan starting at $9 a month on pricing), and there is no free tier beyond the 14-day trial.

Common problems

  • @ChatGPT doesn't respond when mentioned in Slack: First confirm that ChatGPT is enabled for that channel and that you mentioned @ChatGPT explicitly. If your workspace used @Codex before, existing installs keep working for a while, and a Slack admin still has to approve the new permissions for @ChatGPT.
  • @ChatGPT says it can't find the repository or environment: Make sure the Codex Cloud environment isn't set to Only me. Open the environment's Privacy settings and set Who can use to your workspace, and spell out the repository name in your Slack prompt.
  • A teammate replied in a @ChatGPT thread and lost the context: In the official app, each task belongs to the requester's account. Either have the original requester send the follow-up prompt, or summarize the issue, expected outcome, and repository in the new teammate's message so their task has enough context.
  • Your homemade bot replies three times to the same Slack message: Your server took longer than three seconds to return HTTP 200 to Slack's event webhook. Acknowledge the webhook first, then call codex exec or the SDK asynchronously.
  • Merlin says no runtime is available: The engineer laptop hosting the Codex runtime is asleep, offline, or still set to private. Either wake the machine, make the runtime public from its runtime page, or install the Merlin CLI daemon on an always-on desktop or dev box.

FAQ

If everyone who needs to ask a question already has a ChatGPT Business seat and your repos work cleanly in a cloud sandbox, the official @ChatGPT app is the cleanest fit. If you want a single-channel experiment billed by the token, a Socket Mode bot on an API key gets you there. And if your engineers already have Codex dialed in locally and you want the rest of the company asking in Slack and Linear, run Codex on your teammates' machines (Merlin is one way to do it).

Try Merlin for free

Hand work to Claude Code and Codex from Slack, Linear and GitHub. Free for 14 days, no API keys, no card required.

Get started for free

Keep reading