Spaces

A space is a set of people and the capabilities they can reach. Every integration, skill and agent lives in exactly one space, and by default everyone in that space can use the integrations and skills that live there.

Agents are the exception, and it is deliberate: an agent is reachable by its author and by whoever is named on it, never by a space's members as a whole. Administrators manage every agent, but use one in their AI chats only as its author or when named on it. Service accounts are the other one, in a different way. A service account reaches only what is granted to it by name, so adding one to a space gives it no integration, skill or agent. What membership does give it is the space's memory: a service account in a space can recall and write what that space remembers.

Spaces replace groups. A group was a list of people you could point a grant at; a space is a place things live, so most of the time you decide access by choosing where something belongs rather than by naming people one at a time.

The one rule to hold on to

Everything has a home, and its home decides who reaches it by default.

There is no "no space" state. A skill, an integration or an agent created without a space is homed in the Organization space, which every member of your organization is in. For a skill or an integration that is what "available to everyone" means. An agent homed there is still reached only by its author and by whoever is named on it: its home decides where it lives, not who may use it.

The Organization space

Every organization has one, it is called Organization, and you cannot delete, rename, pause or add members to it. Every member of your organization is in it, always.

It is why the Organization space is not like any other space in airlock. Homed there, a skill or an integration has no restriction at all (an agent still follows its own rule above). Any other space is a named group of people who are not everybody.

Open and private spaces

Each space you create is open or private. The type decides who can see the space and how people get in. It does not change what the space can reach.

A new space is private unless you choose open. A space that was closed before open spaces existed shows as private now.

Use open for a space anyone may opt into: a community of practice, a project people drift in and out of. Use private for a space whose contents should stay with its members: an executive team, a legal matter, a security investigation.

Administrators

Administrators see every space, private ones included, and manage every space from the Control Room. Seeing a space is not using it. An administrator uses a space's integrations, skills and agents in their AI chats only after joining it, exactly like everyone else.

An administrator can join any space. Joining a private space is recorded in the audit trail as an admin joining a private space, so the act of taking access is visible afterwards. The space's page tells an administrator who is not a member, with a Join button.

In a private space an administrator who has not joined cannot open what lives there. They can still see which tools and skills are used in it. Approvals are the one place an administrator acts inside a private space without joining it: a request routed to a private space's members can be decided by an administrator too, and that decision is in the audit trail like any other.

One limit to know before you rely on it: a private space withholds the integrations and skills that live in it. Lending one of its skills to another space makes that skill reachable from there. It does not withhold an agent. An agent follows its own rule wherever it is homed, so a private space is not the control to reach for if the thing you want to keep to its members is an agent.

Joining and leaving

Anyone in the organization can join an open space themselves. Open spaces you are not in are listed under Spaces you can join in Settings → Account, each with a Join button. You join as a member.

Anyone in a space can leave it with Leave space on the space's page. Leaving an open space is easy to undo: join it again. Leaving a private space means an owner has to invite you back. The last owner of a space cannot leave it; an administrator appoints another owner first. You cannot leave the Organization space.

An open space cannot approve

An approval rule cannot route requests to an open space. Anyone could join it and approve their own request. A rule can name a private space, the Organization space or a person. A space that an approval rule names cannot be made open until the rule stops naming it.

Exclusive

Exclusive is a setting on any space, off by default. When an exclusive space is the space you work in, your AI only uses that space and Organization: nothing from your other spaces is reachable. An exclusive space cannot be used while another space is the space you work in.

Use it for work that must not mix with anything else: a client engagement, a matter under an NDA. Owners and administrators turn it on or off when they edit the space. The Organization space is never exclusive.

Working in a space from your AI app

You connect your AI app to airlock once, at https://mcp.air-lock.ai/org/<your-org>. There is no separate connection per space. Inside that one connection, you work in one space at a time: the space you work in.

How the space you work in is set

Your AI tells airlock which space you work in by naming the space's key on a call. Say it in the chat ("I'm working on Marketing"), or put it in your project's instructions so every chat starts there:

Airlock remembers the space for that AI connection until your AI names another one. It forgets it after 30 days without use. Your AI can ask airlock for your spaces and their keys at any time.

If nothing names a space yet:

• You are in one space besides Organization: airlock uses that one.
• You are in no space besides Organization: airlock uses Organization.
• You are in several: airlock asks your AI which space you are working in before it saves anything to memory.

Until a space is set, using one of your private spaces does not ask first. The approvals below start once you work in a space.

What changes when you work in a space

• You can still use everything in all your spaces. The space you work in comes first in search results.
• Organization is always there. Its integrations, skills and agents work from any space, and its memory can be read from any space.
• Memory is saved to the space you work in, and only there. The memory of Organization and of your other spaces can be read, not changed. To save to Organization memory, work in Organization.
• Integrations are not read-only like memory. Sending an email or posting to Slack works from any space whose integrations you can use.
• Private spaces ask first. See below.
• Exclusive spaces stay separate. While you work in an exclusive space, your AI only uses that space and Organization. An exclusive space cannot be used while you work in another space.

Using a private space while you work in another one

If your AI needs something from a private space you are in (an integration, a skill, an agent, or its memory) while you work in another space, airlock asks you first. Your AI tells you, with a link: "Using HR while you work in Marketing needs your approval." You approve it in the Control Room from that link (some AI apps also show it as an approval card), and choose how long it stays open:

• For 1 hour (the default)
• For today, until midnight on your clock

Then ask your AI to try again. Airlock does not repeat the call for you.

Only you can approve your own request. An administrator cannot approve it for you, because it opens a space in your name. Airlock sends no email and no Slack or Teams message for it.

The spaces you opened are listed in Settings → Account under Spaces opened in your AI apps, with the AI app and when each one ends. Revoke closes one straight away. Leaving the space, or the space being paused or archived, closes it too.

Some calls also need someone's approval under your organization's tool policy. A call like that made in a space you opened runs only while that space is still open: if the opening ends or you revoke it before the call is approved, it does not run.

Opening a private space lets your AI use it and read its memory. It does not let your AI save to that space's memory: memory still goes to the space you work in.

Leaving an exclusive space

Switching away from an exclusive space asks for your confirmation the same way. The request reads "Leave Acme engagement and switch to Marketing". Once you approve it, Marketing is the space you work in. An exclusive space can never be opened from another space, not even with an approval.

Memory suggestions

After your AI uses a tool, airlock may add one short line suggesting it save what is worth remembering to your space's memory. Your AI decides whether to do it. Airlock never saves memory on its own. A space with memory set to None gets no suggestions.

Agents

• An agent running on its own, for example in Slack under its own service account, reaches only what it was granted. It has no space you work in and never asks to open another space.
• An agent you run in your chat, for example an agent definition you start in Claude Code, works with your access. It reaches your spaces, follows the space you work in, and asks you before opening a private space.

Examples

A day in Marketing in Claude.ai. Your Marketing project's instructions say "work in the Marketing space". You ask for a campaign brief for Acme. Your AI finds Marketing's campaign skill and HubSpot first, and reads what Marketing and Organization memory know about Acme. You say "remember that Acme prefers email": it is saved in Marketing. You ask it to post the brief in #general: it posts.

Sales numbers while you work in Marketing. You are also in Sales, an open space. Your AI uses Sales' HubSpot to look up the Acme deal, no questions asked. You say "remember the deal size": it is saved in Marketing, the space you work in, not in Sales.

Something from HR while you work in Marketing. HR is private and you are in it. Your AI needs the leave calendar. It tells you that using HR needs your approval and gives you the link. You approve for 1 hour and ask again: it works, and for the next hour HR stays open on that AI connection. Anything your AI saves still goes to Marketing. The approval is listed under Settings → Account, where you can revoke it.

An exclusive client space. Acme engagement is private and exclusive. While you work in it, your AI sees only Acme engagement and Organization. It cannot reach Marketing or Sales, and Organization memory can be read, not changed. When you ask to switch to Marketing, airlock asks you to confirm leaving Acme engagement.

Joining an open space. In Settings → Account, under Spaces you can join, you join Product feedback. You are now a member and can work in it.

A private space you are not in. Board is private and you are not a member. You do not see it anywhere. An administrator sees it and can join it, and that join is recorded.

A Slack agent that lives in Marketing. It answers questions with what it was granted, and saves to the memory of the spaces it is in. It cannot reach Sales or HR unless it was granted something there, and it never asks to open another space.

An agent you run in Claude Code. You start the Campaign agent. It works in Marketing, with your access: if you are in Sales, it can use Sales too.

No space named. You are in Marketing only: airlock uses Marketing. You are in Marketing and Sales and nothing names a space yet: when your AI wants to save something, airlock asks which space you are working in.

Two chats on one AI app. Claude Code keeps one connection to airlock for all its chats. If one chat works in Marketing and another in HR, have each chat name its space in its instructions, so neither picks up the other's.

Memory off. Marketing's memory is set to None. Your AI gets no memory suggestions there, and anything it tries to save is refused with that reason. Nothing is saved to Organization instead.

Removed from a space. An owner removes you from HR while you had it open from Marketing. From your AI's next call, HR is no longer available and the opening is gone.

Owners

A space you create in airlock starts with you as its owner. Ownership is not guaranteed to persist: an administrator can remove the last owner, and a space that arrives from your identity provider has no owner until somebody sets one. A space with no owner still works, and everything below is then an administrator's alone.

An owner manages their own space without being an organization administrator:

• Add and remove its members
• Change its name and purpose
• Turn Exclusive on or off
• Lend a skill or an agent homed in it to another space
• Set the tool policy of an integration that lives in it, on the integration's Tools tab: which tools are allowed, which need approval, and who approves

An owner's approvers can only be themselves or a space they own. Your organization's own tool policy still applies on top of what an owner sets, and an integration that lives in the Organization space has no owner, so its policy stays an administrator's.

Owners are members with these extra rights. Everything else stays an administrator's: whether the space is open or private, pausing, archiving, deleting, the organization's tool policy, and every other kind of grant.

Lending

A skill or an agent lives in one space. Lending makes it reachable from another one without moving it. Revoking a lend touches nothing else: the resource stays where it lives, and every other lend stands.

Integrations are not lent. When another team needs an integration, add it in that team's space as well. Each one then has its own access and its own tool policy, so nothing one team changes reaches the other.

An administrator lends from the Spaces tab on the skill or agent, which also lists every space it currently reaches and revokes any of them. An owner's lending runs over the API today; the tab is an administrator's screen.

Moving a resource to another space

An administrator can move a resource's home from three places:

• the bottom of its Spaces tab, where the home is named, with Move home;
• a space's Capabilities tab, with Move on anything homed there, or by dragging its row onto a space under Your spaces in the sidebar. To move something from Organization into a space, open the Organization space and drag it;
• Add capability on a space, with Move here.

A move is deliberately blunt: everyone in the old space loses it and everyone in the new one gains it, in one step. So before you confirm, airlock shows who that is:

• Loses access: the people and service accounts who can use it now and cannot after the move.
• Gains access: those who cannot use it now and can after.
• How many keep using it.

The lists are about using it in AI chats, worked out with the same rule airlock applies to every call, so lends, exceptions and private spaces are taken into account. Administrators manage it wherever it lives, so they appear on a list only when their own use changes. A personal skill is only ever its owner's; moving it out of a paused or archived home restores the owner's access.

Two things a move does not do, and it says so before you confirm:

• Lends stay. If the skill was lent to Marketing, it still is. On a skill's Spaces tab, you can also remove a lend to its new home because that lend becomes redundant. An agent's lend can still grant access in its new home, so it is kept.
• Exceptions stay. A narrowing inside the old space is kept rather than deleted, so moving the resource back restores it.

A move into a paused space is refused, and so is lending anything to one, because nothing in a paused space can be used by anyone, and a grant written while it is paused would come alive the moment it is unpaused.

Exceptions: narrowing inside a space

Sometimes everyone in a space should not reach one particular thing. An exception narrows a resource inside one space to named people, and it is the same access picker you already use, on the resource's Access tab.

Exceptions are rare by design. If you find yourself writing many of them in one space, the space is probably drawn around the wrong set of people.

Pausing a space

Pausing stops everything homed in or lent to a space from being used, by everyone, administrators included. Tool calls are refused, held approvals do not run, and its skills and agents are not served.

Nothing is deleted. Membership, grants and settings stay exactly as they were, and unpausing restores all of it. It is the control to reach for when something is wrong and you want it to stop now rather than to work out what to unpick.

Archiving and deleting

Archiving removes every member and stops anything in the space being reached. Its grants stay stored so an administrator can review them.

Deleting removes the roster, every grant naming the space, every approval rule that routed to it, and the space itself, and drops it from any AI client's allowed spaces (a client that allowed only this space is left allowing none until an administrator widens it). Memory written for the space is kept but readable by nobody, and stays that way after the delete. It is refused while an integration, a skill or an agent is still homed there: a resource needs a home, so move those somewhere else first.

Spaces that come from your identity provider

If you use SCIM, a group in your identity provider is a space in airlock, marked as directory-managed.

Its roster belongs to the directory: adding or removing a person in airlock would be undone by the next sync, so airlock refuses those edits and says why. What stays yours is everything else, including its owners and whether to delete it. When the group is deleted in the directory, the space is archived and its roster, owners included, is removed with it; deleting the archived space in airlock is what lets the directory provision the group again.

Memory per space

A memory server remembers for the space it is homed in. Each space chooses where that memory lives on its Memory tab, and every choice comes down to one instruction your agents are told. Administrators set it.

The Organization space is always shared with the organization, because every member is already in it.

A few things worth knowing:

• The preset decides where new memory goes, not what already exists. Switching a space from shared to kept in this space does not hide what was written while it was shared, and switching to None deletes nothing.
• An agent can still choose a narrower audience. A write may name another audience the caller already belongs to, such as only themselves. It can never name a space the caller is not in, and through a server whose space keeps its memory in the space it cannot name the organization either: nothing written there leaves the space.
• The instruction travels with the server. An agent that discovers a memory server's tools reads the space's instruction at the top of every memory tool's description.
• Archiving or pausing a space seals its memory. Nobody can read or write it through any memory server or the memory view, administrators included. Unpausing restores it exactly as it was. An archived space's memory is kept, not deleted.
• Two spaces can remember the same name separately. An entity called "Acme" in Marketing and one in Finance stay two entities, each with its own facts.

The Memory tab also shows how many entities the chosen preset holds and how many people wrote them, so you can see what a switch would change before you make it.

Finding things by space

Every row is badged with where it lives, a skill or an agent also with where it is lent, and the Integrations list carries a Space dropdown that narrows it to the spaces you choose:

Filtering the integrations on a space keeps the ones homed there. The badge for the space you filtered on is dropped from every row. The dropdown is multi-select, because comparing two spaces is a real question. It always tells you how many rows it is hiding, with one click back to everything, so a filtered list can never be mistaken for a shrinking library. To see everything a space holds across integrations, skills and agents, open the space itself from the sidebar.

Your own spaces are listed in the sidebar. Each entry opens that space: who is in it and what is homed in or lent to it, and for administrators also where its memory lives. The Integrations list is the one list with a Space dropdown, for narrowing it to the spaces you choose.

What happened to groups

Your groups became spaces when you upgraded, keeping their names and their members, and everything they could reach stayed reachable. Two things changed on the day:

• Everything that existed before spaces was homed in the Organization space, so it carries the Organization badge and nobody gained or lost access in the migration. A team that never creates a space sees nothing new to manage.
• The Groups tab in Settings is now the Spaces tab.

If you had a group only to name it in grants, it still works that way: a space can be a grant target exactly as a group was. What is new is the option to stop naming it, and to home the resource in that space instead.

Common setups

A team with its own tools. Create a private space, add the team, and home their integrations and skills there. Nobody outside the team sees them, and onboarding is one membership change.

One integration two teams need. Add it in each team's space. Each team then has its own access and tool policy for it.

Something only some of a team should touch. Home it in the team's space, then add an exception on its Access tab naming the people who may use it.

A space nobody but its members may read, administrators included. Make it private. An administrator who needs in has to join, and the join is recorded.

A space anyone can opt into. Make it open. People find it under Spaces you can join and add themselves. Do not name it as an approver; pick a private space for that.

Client work that must stay separate. Make the space private and exclusive. While someone works in it, their AI reaches that space and Organization, and nothing else.

Gotchas

• A move is not a lend. Moving takes a resource away from everyone in the old space. If you meant "and also this team", lend it, or for an integration, add it in that team's space too.
• Pausing binds administrators too. That is the point, and it is why pause is safe to reach for. If an administrator says a paused space's integration still works for them, something else is wrong.
• Archiving or deleting a space is refused while anything lives in it. Move its contents first; the message names what is in the way. An archive cannot be undone, so it is checked the same way a delete is.
• A private space is not a secret space. Its name is visible to administrators in the registry. What it withholds is access to what lives inside it.
• Administrators are not in every space. An administrator manages every space but uses one in their AI chats only after joining it. If an administrator says a space's integration is missing from their AI client, check that they are a member.
• An open space never approves. Anyone can join it, so an approval rule cannot route to it, and a space a rule names cannot be opened.
• One AI connection, one space at a time. Chats that share a connection share the space you work in. Name the space in each chat's instructions when they differ.
• Opening a private space is yours alone. Nobody else, administrators included, can approve it for you, and it never lets your AI save to that space's memory.