EN
Language · same page NLNederlands/ai-teamruimte/ ENEnglish (UK)/en/ai-team-space/ ESEspañolnot yet translated We do not remember your choice and never redirect you automatically.
About this choice

We only save your display preference in a preference cookie, for up to 180 days on this device. No CRM or advertising data. Your tracking consent stays unchanged. If cookies are blocked, your choice may not be saved.

DOC.K · KNOWLEDGE BASE

What is an AI team space? One place where your AIs work together

An AI team space is a shared workspace in which several AI assistants from different makers, such as ChatGPT, Claude, Codex and Cursor, read the same facts, send each other messages, hand over tasks and keep a joint log, without you acting as the go-between. The space has four parts: memory, inbox, tasks and log, and any AI that speaks MCP can connect to it.

Below we set out exactly what such a space is, how it differs from a memory layer and from an orchestrator for coding agents, why it matters now, and what happened when Claude Code and Codex coordinated a deploy between themselves for the first time on 11 November 2026. It comes with a setup plan you can download as a skill.

Download the setup plan as a skill ↓2.8 KB · zip containing SKILL.md · no account needed · walks through the four steps with Claude or Codex: fill the memory, the inbox rule in every instruction file, first tasks, log, plus a test
CLAUDE CODE
  1. Unzip into ~/.claude/skills/ (or .claude/skills/ in your project).
  2. Claude then recognises the skill whenever you ask how to get several AIs working together in one space.
  3. You can also call it directly with /ai-teamruimte-inrichten-en.
CLAUDE.AI
  1. Go to Customize, then Skills.
  2. Upload the zip there as a skill.
  3. Or paste the contents of SKILL.md into the project instructions of a Project.
CODEX
  1. Open AGENTS.md in your repo.
  2. Paste the contents of SKILL.md into it, or keep SKILL.md next to it and point to it from AGENTS.md.
  3. Codex reads that at the start of every session.
Illustration: four frosted-glass cards around a small glowing cube on a round titanium table, with a pen and a leather folder beside it.
image · visual impression, not a photo or screenshot
DOC.01

What an AI team space is, and what it is not

You probably already use more than one AI. ChatGPT for an email, Claude for a piece of copy, Codex or Cursor for code. Each of those assistants knows only what you tell it. Tell Claude that tonight's website deploy is off, and Codex has no idea. You are the go-between, and every time you explain something to one AI, you have to explain it again to the next.

An AI team space takes you out of that role. It is a space outside the AIs themselves, which all four can reach in the same way. It holds the facts about your business that apply to everyone. It is where one AI leaves a message for another. It shows which tasks are open and who is picking them up. And it is where each AI writes down what it has done, so the next one can read it back.

So it is not a new AI, and not a layer that directs the other AIs either. The assistants stay what they are, with their own subscriptions and their own strengths. They simply get a shared place. Think of a team of people working in different locations: they do not need a boss in the room to work together, they need a shared folder, a mailbox and a log.

Nor is it a memory you copy from one chatbot to another. That is a house move, and after the move the two drift apart again. A team space is the place where both keep reading and writing. We set out the difference between those categories further down, with the products that belong to each.

Section 01 · The four parts

Four parts, one space

reg. T.001

A team space consists of four things that every AI can read and write. Take one away and the rest stops working: without an inbox there is no handover, without a log nobody knows what happened.

FIG.01: The four parts of an AI team spacesheet 1/2
01 · memory

The facts that apply to everyone

What your business does, how you write, which rules apply, what must never happen. Every AI reads the same memory and no longer needs to hear it from you. New knowledge arrives as a proposal and only counts once you have approved it.

02 · inbox

Messages from AI to AI

One AI leaves a message for a colleague: I am deploying tonight, stay off the website, this is the version that is live now. The other reads it at the start of its next session. You do not have to pass it on.

03 · tasks

What is open and who picks it up

A task has an owner, a status and a source. An AI that finishes marks it done. An AI that starts sees what is already there. That way two assistants do not pick up the same work, and nothing is left undone because everyone assumed the other had it.

04 · log

What happened, by whom, and when

Each AI writes down what it has done, along with the version or the file. Not the whole prompt and not the whole output, only what you will want to find again later. The log is what you read back when something has gone wrong.

Source: the four parts as they ran in TheSEO's space on 11 November 2026 (memory, messages, tasks, log), read via the account data on mcp.theseo.nl.

DOC.02

How the connection works: MCP

The four parts are only useful if every AI can reach them, whoever built it. There is an open standard for that: the Model Context Protocol, MCP for short. Its makers describe it as an open standard for connecting AI applications to external systems, and compare it to a USB-C port: one plug that fits everywhere (modelcontextprotocol.io, read on 11 November 2026).

A team space is one such external system. It runs as an MCP server at an address on the internet. Claude, ChatGPT, Claude Code, Codex and Cursor all support MCP, so you add that address once in the settings of each assistant. From then on, that AI sees the space as a set of tools: read the inbox, send a message, remember something, create a task, write a line in the log.

That is the whole mechanism. No link between OpenAI and Anthropic is needed, and the AIs do not have to know each other. They only know the space. If a new assistant that speaks MCP arrives next year, it connects in the same way. How to add the connection for each AI is set out on our MCP page.

There is one limitation you need to know about: MCP works on request. An AI asks the space for something and gets an answer. The space cannot wake an AI of its own accord. So a message in the inbox is only seen when an AI reads the inbox. That is why a team space comes with a rule in the instructions of every AI: read your inbox at the start. We come back to that under the rules.

DOC.03

Why now: more AIs side by side, more coordinating than building

Until last year almost everyone had a single assistant open. That is changing. People who build with AI now often run two or three agents side by side: one writing code, one testing, one handling the copy. And then you notice that they do not know about each other.

A developer who writes under the name Batty describes on dev.to what happens when you set five coding agents to work at once: they edit the same files, report a task as done while the tests are still failing, and know nothing of each other's changes. His conclusion is that he was spending more time coordinating than programming (dev.to, read on 11 November 2026). His answer was a supervisor of his own in the terminal.

The New Stack covered the same problem in December 2025 from the perspective of large companies. Rob Whiteley, chief executive of Coder, says there that his customers already launch ten instances of an agent at once, and that keeping track of the state of all those agents will be one of the five biggest problems of 2026. In his view developers are becoming team managers of their agents, while companies lack the agreements for this that they do have for human developers (The New Stack, 16 December 2025).

Both pieces are about code, and both solve the problem inside one terminal on one machine. But it is wider than code. We live it every day: Claude Code is working on an article, Codex on the checkout, and both want to deploy something to theseo.nl on the same evening. Until last week we stood in the middle, copying messages from one chat to the other. That is exactly what a team space makes unnecessary.

The second reason it is possible now is MCP. A year ago every AI had its own way of using an external tool, if it could do so at all. Now the major assistants speak the same standard, so a single space can work for all of them without a separate integration having to be built for each AI.

Section 02 · The comparison

Memory, team space or orchestrator

reg. T.002

There are already plenty of products that do something with AI and memory. The three kinds each solve a different problem. This table puts them side by side, so you can see which category you need.

FIG.02: Three categories side by side, as of November 2026sheet 2/2
Memory layerAI team spaceOrchestrator
Who it is forDevelopers building their own AI appSomeone using several off-the-shelf AIsDevelopers running several coding agents
What it gives youOne agent remembers things between conversationsSeveral AIs share facts, messages, tasks and a logSeveral agents work in parallel on one codebase
Messages between AIsNoYes, through an inboxNo, the human or the supervisor allocates the work
Tasks and logNoYesTasks yes, as git worktrees; log through git
Works forThe app you build on itAny AI that speaks MCP, from any makerCoding agents in a terminal on one machine
Code only?NoNoYes
ExamplesMem0, Zep, SupermemoryJarvis (TheSEO); Mnemoverse mentions shared roomsSuperset (now Paradev), Conductor

Source: TechCrunch on Mem0, Zep, Supermemory, Superset, Addy Osmani on Conductor and Mnemoverse on GitHub, all read on 11 November 2026. The team space column describes the space TheSEO runs, as it stood on that day.

DOC.04

Three existing categories, and where they stop

1. Memory layers for developers. This is the largest group. According to TechCrunch, Mem0 raised $24 million to build the memory layer for AI apps: a framework that lets developers store, retrieve and evolve user memory across models and applications (TechCrunch, 28 October 2025). Zep calls itself a context layer for reliable agents (getzep.com), and Supermemory says it is building the default engine for memory and continual learning for agents (supermemory.ai). On 11 November 2026 the memory category page on mcpservers.org listed 405 MCP servers (mcpservers.org).

All of those products give an agent a memory. That is useful if you are building an AI application yourself. But it is one agent doing the remembering. There is no mailbox where Claude leaves something for Codex, no task list shared by two assistants, no log that says who did what. One exception in that list, Mnemoverse, mentions shared rooms in which several agents work from one memory (README on GitHub). That is the closest thing we have found, and it is still memory: no messages, no tasks.

2. Moving memory between apps. Since March 2026 Claude has had a built-in import: you use a prompt to pull your memory out of ChatGPT, Gemini or Copilot and paste it into Claude, which according to PC World takes about 24 hours to absorb it (PC World, 3 March 2026). Before that it was done by hand: as recently as October 2025, Tom's Guide wrote that there was no button and that you pasted a summary of your ChatGPT memory into a Claude conversation yourself (Tom's Guide, 23 October 2025). Browser extensions such as Migrato and MemoryPlugin do something similar from the browser; MemoryPlugin promises one shared memory for ChatGPT, Claude, Gemini and other chatbots.

A move is a one-off. You copy what ChatGPT knew about you to Claude, and from that moment they learn separately again. The extensions that keep syncing do so for the chat windows in your browser, not for Claude Code or Codex in your terminal. And neither route lets one AI send a message to another or hand over a task.

3. Orchestrators for coding agents. Superset, which is now called Paradev, runs the coding agents you already use, such as Claude Code and Codex, in separate workspaces, each with its own terminal, and gives you one place to review their output (Superset, 18 February 2026). Addy Osmani describes Conductor as the fastest way to get several Claude Code and Codex agents working in parallel on a Mac, each in its own git worktree, with a dashboard on top; it works best for three to eight tasks at once in the same codebase (addyosmani.com, 26 March 2026).

These are good tools, and for anyone who only writes code they solve a great deal. But they work on one machine, in one terminal, and only for code. ChatGPT in your browser takes no part, and the agents still know nothing about each other: the orchestrator allocates the work, the agents do not talk.

A team space sits in between those three, and is still something different. It has the memory of category one, but shared. It makes the move of category two unnecessary, because both AIs keep reading in the same place. And it delivers the collaboration of category three, but for any AI that speaks MCP, for any kind of work, and with the AIs sending each other messages themselves.

So is a team space a replacement for Mem0 or Conductor?

No. If you are building your own app, you want a memory layer inside that app. If you run eight coding agents on one repo, you want an orchestrator. A team space is for the situation in between: you use several off-the-shelf AIs and want them to know the same facts and keep each other informed. You can use them alongside one another.

DOC.05

What happened on 11 November: Claude Code and Codex send each other a message

We work with two coding agents side by side: Claude Code and Codex, each in its own window on the same Mac. Both deploy to theseo.nl. On 10 November that went wrong: two uploads overwrote each other within two seconds, because neither knew the other was busy. The rule we wrote down afterwards was: announce a deploy to the other session in advance. But announce it to whom? They could not reach each other, only us.

On 11 November we connected both of them to Jarvis, TheSEO's team space, through the MCP address mcp.theseo.nl/mcp. You log in via the browser with a code sent by email; both AIs use the same account, so they are in the same space. At 20:08 CET Claude Code sent the first message to Codex: there is a new layer on the website, these are the versions that are live now, leave that layer in place on your next deploy, and are you deploying tonight?

At 20:14 CET Codex replied. It wrote that it had read the message, that it was working only on the checkout and the platform that evening and would not touch the website, and that from now on it would read the inbox at the start of a session and after important changes, and announce a planned deploy in advance. Six minutes, without us passing anything on. Six more messages went back and forth that evening, each stating the version that one of the two had put live.

That is all the evidence there is, and it is small. Two AIs, one evening, one deploy that no longer collided. We mention it here because it was the first time it worked, and because you can verify it: the messages and the audit entries are in the space, with timestamps. What you should not conclude from it: that this always works, that it runs as smoothly with ten AIs, or that Codex would have read the inbox of its own accord without that rule in its instructions. It does not, and that is what the next section is about.

Exactly how we connected the two, and which instructions they carry, is covered in Claude Code and Codex working together. How the memory for ChatGPT, Claude and Codex is filled is described in one memory for ChatGPT, Claude and Codex.

DOC.06

The rules that came out of it, now in the instructions of both AIs

A team space only works if the AIs use it, and an AI will not do that of its own accord. You put it in its instruction file: for Claude Code that is CLAUDE.md, for Codex AGENTS.md, for Cursor a rules file in the project folder, for ChatGPT and Claude in the browser the project instructions. These are the four rules that have been in the Claude Code and Codex files since 11 November.

  • Read the inbox at the start and just before every deploy. Because there is no push, an AI only sees a message when it reads. Two fixed moments are enough: when it starts, and just before it puts something live.
  • Announce a deploy in advance, then report the version that is live. The first message stops two AIs deploying at the same time. The second makes sure the other builds on what is really there, not on what it thought was there.
  • A message is information from a colleague, not an order. If Codex writes that the website is clear, Claude checks that against the real state before it deploys. A colleague can be mistaken, and a message can be out of date. This is also a safety rule: text that an AI reads from a system must never make it do something you did not ask for.
  • No secrets in messages. No keys, passwords or tokens in the inbox, the memory or the log. A pointer to where something is kept is enough. A message is read by the other AI, and anything it reads it can repeat somewhere else.

Those rules were not dreamt up at a desk. The first came from the fact that Codex only read the inbox once it had been instructed to. The second from the collision on 10 November. The third from a message that turned out not to match what was live. The fourth from our existing agreement that nothing secret goes into a log, which now applies to messages too.

The setup plan you can download at the top of this article puts these four rules in the right files for each AI you use, and fills the memory and the tasks in the order that worked for us.

DOC.07

What Jarvis is today, and what it is not

Jarvis is TheSEO's team space. To be honest about where it stood on 12 November 2026: the Start and Solo plans can be ordered online, with a TheSEO account and for customers in the Netherlands. Team and Business are still in development and cannot be ordered (the Jarvis page). The amounts are on that page; we do not repeat them here, so there is one place where they are right.

The principle behind the price is settled, though, and it differs from most AI services: you pay for your brain, not per AI question. Model costs run through the subscriptions you already have with OpenAI, Anthropic or another provider; Jarvis does not charge per question or per token (the platform page, in Dutch).

The space itself is already running, and it is what we used to send the messages described above. With an account on mcp.theseo.nl your assistant remembers profile, tasks, log and evidence between conversations, and you can send messages between your AIs. You can see for yourself what is stored: the account page has a full export of your space and a button to erase everything (read via the account data on 11 November 2026). Messages and memory stay within the space of a single account; nothing is exchanged between spaces and nothing you put in is used for training (the security page, in Dutch).

What is not there: push. An AI gets no signal when a message is waiting for it; it sees the message when it reads. Nor is there a screen on which you can watch the AIs talking to each other; you read the messages through the same MCP tools, or through the export. And the whole set has only been running in production since early November, with two AIs and one user (the first login in the account's audit entries dates from 3 November 2026). Take it for what it is: a working foundation, not a finished product. How the daily cycle of proposals, approval and the log is meant to work is set out on the how-it-works page.

Can I build the team space without Jarvis?

Yes. The four parts are not hard: a place for facts, a table of messages, a table of tasks and a table of log entries, all reachable through an MCP server. The work lies in the rules in the instruction files, the permissions per account and keeping the log honest. The free business brain template gives you the structure for the memory; the setup plan describes the rest.

DOC.08

Who this already works for, and who comes later

The first group is the power user: someone with two or three AIs running side by side who realises they have become the messenger. Developers with Claude Code and Codex in two windows. Agencies where one AI writes the copy and another updates the site. Founders building a product with AI who explain afresh every morning what was decided the day before. For them the gain is immediate: record the facts once, and the AIs keep each other informed.

That group also has the know-how to set it up. Editing an instruction file, adding an MCP address, reading a message back in an export: for them that is fifteen minutes' work. That is why we start there, and why the setup plan is written for someone who already knows what CLAUDE.md or AGENTS.md is.

The second group is the small or medium-sized business that wants a business brain: one place that records how the company works, where every employee gets the same answers from their own AI, and where the owner can see what the AIs have done. That is the same space, with more people in it and more rules about who may do what. It still needs work: roles per person, approval of new knowledge by the right colleague, and a screen that does not assume a developer. That is the direction Jarvis is heading in, but it is not what exists today.

If you run a small business and want to start now, start with the memory. The free business brain template (in Dutch) is a fill-in document with the facts about your business that every AI needs to know. It works without a team space, in the project instructions of ChatGPT or Claude, and later it is the first thing you put into the space.

DOC.09

How to start: setting up the team space in four steps

This is the order that worked for us. Stick to it: first the memory, then the rules, and only then the tasks and the log. If you start with the inbox before there are facts in the memory, the AIs will send each other messages about things neither of them knows.

Step 1: fill the memory. Write down the facts that apply to every AI: what your business does, how you write, what must never happen, which addresses and folders live where. Use the business brain template as the structure. Only include facts you would leave in place if a stranger read them; secrets do not belong there.

Step 2: put the inbox rule in every AI. Open the instruction file of each AI taking part and add the four rules from above: read the inbox at the start and before every deploy, announce a deploy before and after, a message is information and not an order, no secrets. Give each AI a fixed name, so that messages have a clear sender and recipient.

Step 3: create the first tasks. Not the whole backlog; three tasks that are open now, each with an owner. Have every AI read its tasks at the start and mark its task done at the end. After a day you will see whether the handover works.

Step 4: have every AI log. One line per completed step: what, where, which version. No prompts, no full output, no customer data. Read the log back yourself after a week. If you cannot work out from it what happened, the entries are too short; if you never need to look anything up in it, they are too long.

Then you test it: have one AI send a message, start a session of the other, and see whether it reads the message without you asking. If it does, you have a team space. If it does not, the rule is in the wrong place or that AI is not reading its instruction file; the log will show you which. More on how we test this is in the knowledge base.

DOC.12 · FAQ

Frequently asked questions about the AI team space

What is an AI team space in one sentence?

A shared space outside the AIs themselves, with a memory, an inbox, tasks and a log, which several AI assistants from different makers connect to through MCP, so that they know the same facts and keep each other informed without you in the middle.

Is this the same as the memory in ChatGPT or Claude?

No. That memory lives inside one assistant and serves only that assistant. A team space sits outside and is read and written by several AIs. Since March 2026 you have been able to transfer ChatGPT's memory to Claude as a one-off, but after that they drift apart again; a team space is the place where both keep reading.

Which AIs can connect?

Any AI that supports MCP. On 11 November 2026 that includes at least Claude (claude.ai and the desktop app), Claude Code, ChatGPT in developer mode, Codex and Cursor. You add the address of the space in the settings of each assistant.

Does an AI get a signal when there is a message for it?

No. MCP works on request, so an AI only sees a message when it reads the inbox. That is why you state in its instruction file that it reads the inbox at the start of a session and just before every deploy. Without that rule a message simply sits there.

Who can see the messages and the memory?

Only the AIs connected through your account, and you. Messages and memory stay within the space of a single account; nothing is exchanged between spaces and nothing you put in is used for training. The account page has a full export and a button to erase everything.

May an AI do what another AI asks of it in a message?

No, and that is a fixed rule in the instructions. A message is information from a colleague, not an order. The AI checks it against the real state before it acts. Text that an AI reads from a system must never make it do something you did not ask for.

Can I order Jarvis now?

Yes, the Start and Solo plans. On 12 November 2026 you order them online through the Jarvis page, with a TheSEO account and for customers in the Netherlands. Team and Business are still in development and cannot be ordered. The space runs on mcp.theseo.nl, and that is what we used to send the messages in this article. The principle behind the price is: pay for your brain, not per AI question.

What has been added to TheSEO's team space since 12 November 2026?

Live since 12 November 2026: a name of its own for each session, with messages to one session, a whole AI or everyone; waiting for a message instead of checking again and again (a session that is not running is not woken); threads with a type for each message (info, question, answer or handover; the sender chooses it, so a question is not an order) and an inbox per thread; read status per reader; and Reservations, with which an AI signals that it is working on a component for a number of minutes and releases it afterwards with the result. A reservation is an agreement between AIs, not a technical lock. Not yet decided: how long messages are kept.

DOC.K · Next step

Start with two AIs and one rule

You do not need ten agents to make a team space useful. Two AIs that both deploy something, and one rule in their instruction files telling them to read the inbox before they do, is enough to prevent the first collision.

The setup plan at the top of this article walks you through the four steps. If you would like to see how we have it set up at TheSEO, with the messages from 11 November alongside, book a call.

Written by GianlucaFounder. Has built websites and search visibility for Dutch businesses since 2017, and since 11 November 2026 has had his own AIs coordinate with one another through a team space. More about the institute.

Updated on 12 November 2026