How to Run a Team of AI Agents in Slack-Style Chat (2026)

Updated 22 October 2026 · ai agents, claude code, codex, team chat

If you want to know how to run a team of AI agents in Slack-style chat using your own Claude and Codex subscriptions, here is the short answer as of October 2026. Neither Anthropic nor OpenAI documents a ready-made team chat for this. You build a small bridge yourself, keep each agent on its owner's paid account, and settle the licensing question before you write any code.

Almost every tutorial skips that licensing question, so this guide covers it first. After that we get to the build, the costs and the places where a setup like this tends to break.

What a team of AI agents in chat really is

It is a chat channel where a message can be sent to an AI coding agent, and each agent runs on a subscription somebody on the team already pays for. The chat app is only the front door. The agents do the work on machines that are logged into their owners' accounts.

In practice the setup has four parts. The first is the chat space itself: Slack, or any tool with channels, threads and bots. The second is a router bot that reads a mention, decides which agent should get the task and posts the answer back into the thread. The third is the set of agent workers, for example a Claude Code session on one person's machine and a Codex session on another person's. The fourth is a log that records who asked for what, which agent did it and what came back.

The research for this article was checked on 22 October 2026, and it found no vendor documentation for any of the following:

• a Slack integration that runs Claude Code and Codex side by side
• a multi-agent router
• an agent-per-user subscription setup
• a permission model for a design like this

That means every piece of glue in this design is yours to write and yours to maintain. If you have never driven an AI agent from a terminal before, read what Claude Code can do for non programmers first, because this project assumes you are comfortable there.

What your own Claude and Codex subscriptions cover

Your Claude subscription covers Claude Code, and your ChatGPT plan may cover Codex. In both cases, though, the usage limits are shared and vary by plan, so neither one gives you a dedicated pool for agents. Anthropic's pricing page, checked in October 2026, says Claude Code is included with paid Claude plans and does not list a separate Claude Code fee. Anthropic's Claude Code documentation describes it as a coding agent that runs in a terminal.

The detail that matters most for a team chat is on the same Anthropic pricing page. Claude Code shares the account's usage pool with Claude on the web, desktop and mobile. An agent that works through a long task in your chat channel spends the same allowance its owner uses for everyday chats.

For Codex, a third-party documentation site that tracks Codex pricing said in October 2026 that Codex is included in ChatGPT Free, Go, Plus, Pro, Business, Edu and Enterprise plans. The same site said features and limits vary by plan, region and workspace settings. OpenAI's own help page, checked on 22 October 2026, says Codex Cloud is available to eligible Plus, Pro, Business, Enterprise, Healthcare and Education accounts, subject to rollout and workspace settings. The research did not turn up a complete official price and limit table for Codex, so check the plan page inside your own ChatGPT account before you plan around it.

Plan prices side by side

Every price below was published by the source named in the row and was current when checked in October 2026. Prices and plan names change, so re-check each one on the vendor's own page before you buy.

Can you pool personal Claude and ChatGPT subscriptions?

None of the sources we checked answer this, so assume you cannot until the vendor's own terms say you can. The research, checked on 22 October 2026, found nothing that establishes whether a company may safely or contractually pool employees' personal Claude or ChatGPT subscriptions into one shared agent system.

That gap is why this section comes before the build. A personal subscription belongs to the person who pays for it. A shared bot that sends the whole team's requests through one person's login turns a personal plan into shared infrastructure. Nobody in the public sources has confirmed that this is allowed.

Here are three rules that keep you on safer ground while the question stays open.

1. One agent, one owner, one account. Each agent worker runs only on the account of the person who set it up. That person is the only one who sees the login.
2. No shared credentials. Never put one person's Claude or ChatGPT login on a server that the whole team can reach. If the bot needs secrets, store only the chat bot's own token on the router.
3. Read the terms for your plan type. Anthropic sells Team plans and OpenAI sells Business plans. If several people will rely on the same agents every day, compare a workplace plan against personal seats before you commit, and read what each one says about shared and automated use.

There is a data question too. Anything posted in the channel can end up in an agent's context, and that includes client names and pasted emails. Our guide on how far you can trust an AI chatbot with customer data covers the checks to run before client material goes anywhere near a model.

How to run a team of AI agents in Slack-style chat, step by step

The build comes down to giving each agent a job and an owner, putting a router bot between the chat and the agents, and making sure a human approves anything that ships. The steps below describe that design in general terms. They do not describe a documented vendor feature, because the research found none.

1. Give every agent a single, narrow job. For example, one channel could send build tasks to a Claude Code worker and another could send review tasks to a Codex worker. When each agent has a narrow job, it is much easier to see which one went wrong.
2. Run every worker on a machine its owner controls. Each worker should be a separate user account or a separate machine, signed in by the person whose subscription it uses. If something goes wrong, that person can switch it off in seconds.
3. Write the router as the only piece that talks to chat. The router listens for mentions, looks up which worker handles that channel, passes the task along and posts the reply back in the same thread. The research did not cover how each vendor supports running its agent from a script, so read the current Claude Code and Codex docs for the supported way before you write this part.
4. Add an approval gate before anything leaves the sandbox. The agent proposes a change, and a named human approves it in the thread before it is merged, sent or deployed. Anthropic's available documentation does not show built-in approval or audit controls for this kind of team setup, so the gate is your code.
5. Log every request and every reply. Write the asker, the agent, the task and the result to a file the team can read. When an output looks wrong a week later, the log lets you trace it.
6. Tell the channel when an agent hits a limit. Have the router post a clear message when a worker stops because of a usage limit, and give the task to a person or a backup agent. Silent failure is the worst result in a shared channel.

If you already run automations elsewhere, the router is the same kind of glue as an n8n workflow. Our step-by-step guide to sorting Gmail with n8n and Claude shows the habits you would reuse here: test on a copy first, add a human check and log everything.

Where the usage limits bite

The limits bite because agent work and normal chat use the same rolling allowance, and the exact size of that allowance is not published. According to Anthropic's pricing page as of October 2026, paid Claude plans have rolling five-hour limits plus weekly limits. The page does not state exact message, token or task quotas.

Here is what that looks like in a team channel. Suppose a teammate queues a large refactor for their Claude Code agent late in the afternoon. The agent spends that person's shared pool, and the next morning their own Claude chats may already be close to the limit. Nothing is broken, because that is how a shared pool works. It still surprises people who thought of the agent as a separate service.

On the Codex side, the third-party Codex documentation site said in October 2026 that limits vary by plan, region and workspace settings. Two teammates on different ChatGPT plans may therefore get very different amounts of work from the same Codex setup, and you will only find out how different by watching.

There are two practical defences. Put a cap on how many tasks each worker accepts at once, and keep a fallback. If one vendor has an outage or an agent runs out of allowance halfway through a job, the team still needs a way to finish. Our piece on keeping a backup AI for when Claude or ChatGPT goes down covers how to plan that fallback.

Pros and cons of the do-it-yourself agent team

The main advantage is that you reuse money you already spend. The main drawback is that you own every piece of the system, including the parts that fail without any warning. Here is the honest balance.

Pros

• Spend you already have. Claude Code comes with paid Claude plans, per Anthropic's pricing page in October 2026, so there is no extra agent bill for Claude.
• A visible record. Work that happens in a thread shows the whole team who asked, what the agent did and who approved it.
• A second opinion built in. If Claude Code and Codex are both in the channel, you can ask one to review what the other wrote and catch mistakes earlier.
• Owners stay in control. Each person can switch off their own agent without touching anyone else's.

Cons

• No documented integration. As of 22 October 2026, the research found no vendor-documented Slack setup for these agents, so a vendor update can break your glue code.
• The licensing question is open. No source we found confirms that pooling personal subscriptions for a team is allowed.
• The limits are vague and shared. Exact quotas are not published, and agent jobs spend the same pool as the owner's everyday chats.
• Safety is all yours. Approval gates, audit logs and permissions are code you write, test and keep working.
• Uneven Codex access. Codex features depend on each person's plan, region and workspace, so the same task can behave differently from one teammate to the next.

Who should not build this

Skip this project if any of the following describes you. That covers most small businesses, and that is a fair result.

• Teams that handle client data under contract. Wait until a lawyer or the client has approved AI agents touching that data.
• Solo owners who cannot maintain a script. A broken router at a bad moment can cost more time than the agents save.
• Anyone who needs predictable output. Shared pools with unpublished limits make it hard to promise a client a deadline.
• Companies whose staff would bring personal accounts. Sort out the plan terms before anyone connects a personal subscription to a company channel.

What the viral videos promise

Some videos in circulation say you can run your personal subscriptions as a background agent team, build reusable skills and even clone whole apps from a chat channel. The research for this article treated those videos as leads to check, not as proof, and the search results checked on 22 October 2026 did not confirm any of those claims. Treat them as possibilities until you see documentation from the vendor.

FAQ

Does Claude Code cost extra on top of a Claude subscription?

No separate Claude Code fee appeared in the sources checked in October 2026. Anthropic's pricing page says Claude Code is included with paid Claude plans. It also says Claude Code shares the account's usage pool with web, desktop and mobile chats, so the agent spends allowance you would otherwise use yourself.

Can my whole team share one Claude Max account through a chat bot?

The public sources do not say this is allowed, so plan as if it is not. One login used by a whole team through a bot is the kind of pooling no source we checked endorses. The safer path is one account per person, or a workplace plan such as Claude Team, after you read its terms.

Which ChatGPT plans include Codex?

A third-party Codex documentation site said in October 2026 that Codex is included in ChatGPT Free, Go, Plus, Pro, Business, Edu and Enterprise plans, with limits that vary. OpenAI's help page, checked on 22 October 2026, lists the plans eligible for Codex Cloud and says access depends on rollout and workspace settings. Check your own account to see what you actually have.

What happens when an agent hits its usage limit mid-task?

The agent stops until its rolling window resets. According to Anthropic's pricing page as of October 2026, paid Claude plans also have weekly limits on top of that window. Your router should post a clear notice in the thread and send the task to a person or a backup agent, so the team does not wait on a job that has already stopped.

Is there an official Slack app that runs Claude Code and Codex together?

The research checked on 22 October 2026 found no vendor documentation for one. Any setup that puts both agents in one channel today is something you or a third party built. Judge it on its code and its terms, and do not assume either vendor supports it.

Verdict

A team of AI agents in a shared chat is buildable, and it can save money because the Claude side comes with plans you may already pay for. It only makes sense when you follow three rules: one owner per agent, a human approval gate and a written answer on the licensing question. For most small teams, the sensible first step is to keep a single agent in a single channel for a while, see how much of the shared pool it uses, and expand only if the numbers and the terms both hold up.

Related reading

• Claude Code for non programmers: what it can do
• ChatGPT vs Claude for research tasks
• How much you can trust an AI chatbot with customer data

About the author

The Tool Signal editorial desk wrote this guide. The team covers AI tools and automation for freelancers and small businesses, and it names real prices along with the trade-offs. Every price, limit and capability above comes from vendor pages, vendor help docs or named third-party summaries checked on 22 October 2026. Each one is labelled in the list of checked claims at the end of the page, and none of it is presented as our own hands-on test. If you are still deciding how to run a team of AI agents in Slack-style chat using your own Claude and Codex subscriptions, start with the terms and the limits, and build only after both check out.