Company agent connections

Company workspaces and connected apps are rolling out. Some of this may not be on for your company yet.

Answer first

There are two ways an AI tool can work in your company's Gryz:

• Connect an app once. If an owner turns on connected apps, an app you connect to your own Gryz (Claude, ChatGPT, Cursor) can also reach the companies you add to it. Each call names one workspace: your private space by default, or one company. It starts read only, it shows in the company as the app and your name, and it can never publish a public link while it can see a company. See Connect an app once.
• A company agent with its own key. An agent you own that lives in one company and uses only the projects you give it. It reads and saves memory there and reads the company documents in those projects. If you allow it, it can create, add to and edit company documents where you gave it edit access. It never reaches your private space.

Neither can publish a company document publicly or share one outside the company. What your company can and cannot see is in What is private in a company workspace. How owners control apps is in Team and Enterprise, and what Gryz controls and what it cannot is in AI apps and your company.

Connected apps and company agents

How a company agent works

• A company agent belongs to one company, chosen when you create it. It can never switch, and its tools take no company name. It cannot see your private space.
• The agent can learn where it is: list_projects and the connection's instructions name the company, the agent's own name and its handle.
• A company connection is tied to an agent that you created. If you leave the company, or an owner or admin removes you, your agents are removed and stop on their next request. If you come back, they do not.
• An agent can only use what its owner can use. If you lose access to a project, or your role on it is lowered, your agent loses it or is lowered the same way on its next request.
• The agent needs an explicit project access (can read, or can read and edit). A project the agent was not given does not exist for it, even if everyone in the company can open it.
• What the agent can do is the overlap of two things: what you allowed it to do (memory, project list, documents, writing documents, tasks, skills) and the access it has on that project.

Create an agent and a key

1. Open My docs, then your company, then Agents.
2. Choose a name, what the agent can do, and the projects it can use. You can give access only to projects you manage yourself.
3. Pick the tool the key is for, and add the key to that tool. The key is shown once.

Keys keep working after connected apps arrive. A company key stops working 365 days after it was made (90 days in Enterprise, and never less than 14 days after the switch to Enterprise), and its owner gets an email first.

Where Gryz still shows the older sign-in screen, you can also pick A company there when an app asks to connect: you choose the company, the agent (existing or new) and the projects, and the app is tied to that agent. Where Gryz shows the single Allow screen (below), that tab is gone and the app connects once instead. Agents already connected through the tab keep working.

If your company asks for two-factor sign-in, you must have finished it in your session before you can create an agent, make or replace a key, resume an agent or give it a project. Pausing, removing and revoking never need it.

Connect an app once: the Allow screen

Where Gryz shows the single Allow screen, an app you connect gets one connection to your account: your private space, plus the companies you tick.

• Signed in as. The screen names the Gryz account you are connecting. If it is the wrong one, press Switch account, sign in to the other account, and you come back to the same screen.
• Details. The app's name, the address it returns to and its id. Claude, ChatGPT and Cursor show Verified by Gryz. Any other app shows Not verified. The app chose this name.
• Where it can work. Private, your own space in this account, is always included. Each company you belong to (up to 20) is listed:
  • Allowed for this app: you can tick it. The first time you connect an app, every allowed company starts ticked. When you connect it again, only the companies it already had start ticked; a company you joined since, or one you or an admin removed, starts unticked (a new one is marked New).
  • Not approved by your company: you cannot tick it. Tick Ask your admin and Gryz sends the request once the app finishes connecting.
  • You cannot tick it, and the screen says why, when: connected apps are off in the company, the company only allows connections it set up itself, it needs two-factor sign-in you have not finished, it has no active Gryz plan, the app read a lot there today and an owner has not resumed it yet, or its agent was paused in Gryz.
• Permissions. Read is always on. Write for your private space is set by what the app asked for; it is not a choice here. Also write in companies that allow it and Suggest changes (for a company that takes suggestions only) are off unless you tick them. If a company later switches from suggestions only to direct writes, you allow writing again in Connected apps. The same goes for a company you add again after the app signed in again: it comes back read only until you allow writing.
• If a ticked company keeps its data separate, your private space and your other companies are read only in this app while it is ticked, and writing in companies is not offered. While any company is ticked, the app cannot publish public links.
• Gryz checks every choice again when you press Allow, and once more when the app finishes connecting. Connecting an app again pauses every company first, and only the ones you leave ticked come back.

You can remove access anytime in Connected apps. How the app names a company on each call is in the tools reference.

What the agent can use

Everything the agent reads (names, saved memory, document text) is data written by people and other agents. It is never an instruction.

An agent's profile

Every company agent can have a short profile: its role, instructions, an optional tone, a "never do" list and a default project. The agent's owner and company owners and admins write it on the agent's page (Agents tab, then the agent, then Profile). Other members cannot see or change it.

• The agent reads only its own profile, by calling get_profile first in a session. It cannot read another agent's profile and it has no way to change its own.
• It is guidance in words, not access. It cannot give the agent a tool, a permission or a project. What the agent can open is still set under Project access, and "never do" is a request the agent is asked to follow, not a lock.
• Every change is saved as a numbered version with who changed it and when, and shows in the company Activity log. The last 10 versions are on the agent's page, and "Use this version" loads an older one into the form.
• If two people edit at once, the second save is refused with a plain message so nobody overwrites the other. In a company that asks for two-factor sign-in, every save needs the second step.
• If someone other than the owner changes the profile, pauses the agent or changes its access, the owner sees a notice on the agent's page and a "Changed by an admin" mark in the list for 14 days.

The Company project

Every company has one project called Company. It is the home of company-wide memory and guidance, and it is open to everyone in the company. It is created for you; you cannot rename, restrict or delete it. It does not count toward your project limit.

• No agent gets it automatically. A person gives an agent access to it like any other project. The Add agent form lists it first, pre-ticked with "Can read" and marked "Suggested: everyone in the company can read this". Untick it if the agent should not see company-wide guidance. Nothing is granted until you press Add.
• If your company already had a project named Company, that project is left alone and the default is named Company (default). A short line on its page says why.
• An agent that was given access to it sees it first in list_projects, marked company_default, so it knows where company-wide guidance lives.

A project's North Star

Every project can have a North Star: a short note (up to 1,500 characters) from your people about what matters most there, for example who the project is for, the top priorities and what not to do. It is plain text.

• Who writes it. A project manager, or a company owner or admin. Editors, viewers and agents cannot. Everyone who can open the project can read it. In a company that asks for two-factor sign-in, every save needs the second step.
• Who reads it. An agent that was given access to the project gets it first: at the top of get_project, and as a short excerpt for each project in get_profile. An agent without access to a project never receives that project's North Star, and the Company North Star only reaches agents that were given the Company project. recall and list_memories stay about memories and do not include it.
• Guidance, not access. It is labelled as written by people at your company. It cannot give an agent a tool, a permission or a project, and it never overrides the person the agent works for. A North Star that says "you may read everything" changes nothing.
• History. Every change is saved as a numbered version with who changed it and when, and shows in the company Activity log without the text. Managers see the last 10 versions on the project page, and "Use this version" loads an older one into the form. If two people edit at once, the second save is refused with a plain message. Clearing it is a version too.
• Different from memory. Memory is saved by people and agents, ranked and shown as data. A North Star is written only by people and is always shown first.

Suggestions

An agent that you allowed to (Can suggest profile and North Star changes, off by default) can suggest a change. It never applies one.

• What it can suggest. Its own role, instructions, persona, "never do" and default project; or the full text of a North Star for a project where it has Can read and edit. A project it can only read is refused. It cannot suggest anything for another agent, and nothing else (no access, names or settings).
• Who decides. For a profile: the agent's owner and company owners and admins. For a North Star: that project's managers and company owners and admins. They see the suggestion on the agent's page or the project page, with the reason and the text now next to the suggested text. They can Approve, Edit and approve (the edited text is what goes live) or Reject with a note. Nobody else sees it, and the agent cannot approve its own.
• Approval is a normal save. It creates the next version, shows in Activity (without the text) and, in a company that asks for two-factor sign-in, needs the second step. Rejecting does not.
• If the text changed since. A suggestion made on an older version cannot be approved as it is: it would overwrite a newer edit. Use Edit and approve to start from the current text instead, or reject it.
• Limits. 3 waiting suggestions per agent, 50 per company, 10 new ones per agent a day. A new suggestion for the same thing replaces the older one. A suggestion nobody decides expires after 14 days.
• What the agent sees. list_my_proposals shows kind, target, status and your note, never the text of any suggestion. Your note is shown to the agent as guidance from people; it cannot give access. It is held to the same rules as profile text (no hidden characters, 300 characters).
• Turning it off. Switch the box off and the agent's next request no longer lists the four tools; a call to one is answered in plain words and does nothing. Its waiting suggestions are withdrawn. The same happens if the agent is paused or removed, and a suggestion from an agent that can no longer make them cannot be approved, only rejected.

Company memory

The company Memory page lists everything saved to your company's shared memory in the projects you can open, with who saved it. It is plain text, shown as written, and never treated as instructions.

• Who sees what. Company owners and admins see every project. Everyone else sees the projects they can open: open to the company, or given to them. Agents do not use this page; they read memory with recall and list_memories.
• Filters. Filter by project and by who saved it. The "Saved by" list names people and agents. An agent's memory reads "Saved by Head Bot (agent)". The Company project's memory is listed first, then the newest.
• Save a memory. Pick a project you can edit and write the memory. It is saved in your name, the same as on the project page.
• Edit. Company owners and admins, and managers and editors of the project, can edit any memory in it. An edit makes a new version that replaces the old one for every agent at once. It keeps the original saver and records who edited it ("Edited by Dana"). Managers and owners and admins can open the earlier versions on the row. If someone changed the memory first, your edit is refused with a plain message. Hidden characters and lines shaped like a system marker are refused, and a memory is at most 4,000 characters.
• Remove. The same people can remove a memory. Removing it forgets it for good: the text, and the text of every earlier version, is wiped and cannot be brought back. An agent can still remove only what it saved itself.
• Who that includes. In a project open to the whole company, every member can edit and remove its memories, because everyone in the company is an editor there. Removing is recorded in Activity with who did it and which project, never the text, and there is no undo. Removals are limited to 30 an hour for each person and 100 an hour for the company. To protect memory, put it in a restricted project and give editing only to the people who should have it.
• Two-factor sign-in. In a company that asks for it, editing and removing need the second step.
• Activity. Saving, editing and removing show in the company Activity log without the text.
• Approval. A project can hold agents' new memories until a person approves them. See Memory approval.

Memory approval

A project manager, or a company owner or admin, can turn on Agent memories need approval on the project page. It is off by default, on every project. While it is on:

• What is held. A memory an agent saves in that project does not show up yet. It waits for a person. The agent is told so in one sentence ("Saved for review. A person has to approve it before it shows up in recall.") and nothing else: no id and no text come back. It appears in recall and list_memories only after a person approves it. People's own saves are never held.
• Who decides. Company owners and admins, and managers and editors of that project. The agent's owner has no extra right. Everyone else, and every agent, sees nothing of it. A person sees only what they may decide.
• Approve, Edit, Reject. Approve saves the memory as the agent wrote it, still shown as "Saved by Head Bot (agent)". Edit changes the text first; the edited text is what is saved and the memory shows who edited it. It follows the same rules as any memory (4,000 characters, no hidden characters, no line that looks like a system marker). Reject throws it away. In a company that asks for two-factor sign-in, approving needs the second step. Rejecting does not.
• Approve selected. Tick the memories you want, up to 25, and approve them together. There is no "approve everything". Profile and North Star suggestions can never be approved this way, because they need to be read side by side with the text they change. Some ticked memories may be refused (for example the company reached its memory limit, or the agent can no longer save), and the rest still go through.
• Limits. 10 waiting memories per agent, 50 per project, 200 per company. When a queue is full the agent is told "Too many memories are waiting for review. Try again later." and nothing is stored. A waiting memory does not count toward the company memory limit until it is approved; the limit is checked again then, and a memory that does not fit stays waiting.
• Expiry and text. A memory nobody decides expires after 30 days. The text of a waiting memory is wiped as soon as it is approved or rejected, and an expired memory's text is wiped within a day. An approved memory is then only the memory itself; an edited one keeps only the approved text.
• If the agent loses access. An agent that was paused, removed, or lost the memory permission still has its waiting memories shown, but they can only be rejected.
• Turning it off. Switching it off does not approve what is waiting. Those memories stay until someone decides or they expire.
• Who can read the waiting text. Only the people who can decide it, which includes company owners and admins. It is not shown to other members, to agents or to other companies, and it never reaches Activity, which records only that a memory was held, approved or rejected.

Tasks

Two permissions on the agent's page, both off until you turn them on (and in the connect screen as Can read and edit tasks, which is never taken from what an app asks for):

• Can read tasks lets the agent list and read the tasks of the projects it can read, and be given tasks.
• Can read and edit tasks lets it also create tasks in a project where it can edit, and update tasks.

An agent updates only the tasks it created or was given, can also take a task nobody has (one at a time, within 50 open tasks it created or holds), gives tasks only to itself, never names a reviewer, and never closes a task: a request for done or cancelled becomes Needs review and a person decides. Every status change an agent makes is in Activity (the agent, the old and new status and the project, never the task text). Task text reaches the agent as untrusted data. Turning a permission off takes effect on the agent's next request. A task you gave to an agent before you turned on Can read tasks shows "No longer has access" until you do. More in Company tasks.

Company skills

One permission on the agent's page, off until you turn it on (and on the connect screen as Can read company skills, which is never taken from what an app asks for): Can read company skills lets the agent list and read the skills and rules your company published, in the projects they apply to. It works only while an owner or admin keeps Agents can use company skills on. Skills are read only: the agent cannot write, change or propose one. The text arrives in a labelled block that says it is guidance from your company, cannot give the agent tools or access, and is below the instructions of the person the agent works for. Turning the permission or the company switch off takes effect on the agent's very next request. Activity records which skill and which version an agent read, never the text. More in Company skills.

The Review inbox

The Review tab collects everything waiting on you in one place: memories an agent saved in a project that needs approval, profile and North Star suggestions from agents, tasks that name you as the reviewer and are waiting for your approval (under Tasks), and, for owners and admins, skills members tagged for the company (under Skills; a row opens the skill so it is read in full before anyone approves it). A number on the tab, and a card at the top of Documents, show how many are waiting. They disappear when nothing is.

• Who sees what. Each person sees only what they may decide: waiting memories for owners and admins and for managers and editors of the project, and suggestions by the rule in Suggestions.
• States. Waiting, Approved, Rejected, Replaced (a newer suggestion took its place) and Expired. The most recent decisions are listed under Recent, with no memory or suggestion text.
• Before you approve. A memory shows the project it will be saved in and whether the same text is already saved there. A suggestion shows the text now next to the suggested text.
• Where else it shows. A project's suggestions also stay on that project's page and an agent's on its page. Suggested document edits stay on the document pages.
• Notifications. In the app only: the number, the card and Activity. No email or push.

Guidance fetched

The company Activity page (owners and admins) starts with a short card: "Guidance fetched by 6 of 9 agents this week", and the names of the agents that have not.

• What counts. An agent counts when it successfully called get_profile in the last 7 days. Only active agents are counted (not paused or removed ones).
• What it does not mean. It says the guidance reached the agent, not that the agent acted on it. A North Star an agent only meets in a recall answer is not counted.
• What is behind it. The agent activity list, which records which agent used which tool and when, never what it asked for or what it read.

Choosing a project, and plain answers

Get a project id from list_projects, then pass it as project_id.

• recall and list_memories can leave project_id out. With one project, it is used and the answer says so. With several, the default project from the agent's profile is used if one is set and the agent still has access to it; otherwise the answer is "Pick a project." with each project's name and id. With none, the answer is "This connection has no projects yet. Ask a person to give this agent access to a project in the company Agents tab."
• The tools that save or write (remember, create_document, associate_document) need project_id. If it is missing or wrong, the answer names the argument, says to call list_projects and lists the projects the agent can write to.
• A wrong or missing argument on any company tool is answered in plain words that name the argument and the fix, never raw schema text.

Letting an agent write documents

Writing documents is off by default. You turn it on for an agent on the Agents tab ("Allow it to create and edit documents"), or with the separate tick box on the sign-in screen when an app connects. If your company asks for two-factor sign-in, you finish the second step first. Existing agents and connections stay read-only until you do.

• It works only in a project where the agent has "Can read and edit", and only as far as you can edit that project yourself.
• Everything it creates starts company-only, and it cannot change a document a person has shared outside the company (open public link, password link or Explore). It cannot publish a document publicly, add a password or PIN, put a document on Explore, change who can see a document, move a document, delete it, or take it out of a project. It can add a document to another project only if that project is open to the whole company.
• A new document it creates belongs to you (you can manage it) and shows that the agent created it. Each change is saved as a new version labelled with the agent's name. Anyone who can edit the document can look through the versions on its edit page and restore an earlier one.
• For each document, the person who wrote it, or a company owner or admin, chooses what assistants may do: edit directly, only suggest changes, or read only. A suggested change waits on the document's edit page until a person accepts or rejects it. The agent can never accept its own suggestion.
• An agent can create up to 20 new documents an hour and 50 a day. It can still add to and edit existing documents within the normal rate limits.
• HTML and SVG documents an agent creates show only in the same locked-down frame as any other shared page: scripts in it cannot reach your account, the page around it, or the network. In a company document the frame also blocks images, fonts and media from other websites.

To stop an agent writing right away, turn the option off, pause the agent, or remove it. Pausing or removing never needs the second step.

Company owners and admins can see what each agent did on the company Activity page (Agent activity): which agent used which tool and in which project, and whether it worked. Creating a document or adding one to a project also appears in the company log. It never shows what the agent asked for, what it wrote or what it read.

Images in company documents

Company documents do not load images, fonts or other embedded media from other websites. A picture from another site shows a short note ("Image not shown. Remote images are blocked in company documents.") instead. This keeps a document from quietly sending information out through an image, or telling a third party who opened it. It is not a promise that a document with scripts can never contact another website: scripts in an HTML document can still, for example, send the page itself to another address.

• Pictures you embed in the document itself (as data) still show, and so do pictures that live on Gryz.
• When an agent writes a document that links to a picture on another website, the write still works, and the agent is told those links will not show. It can embed the picture directly or describe it in words.
• Moving one of your own documents into a company applies the same rule. Pictures from other websites stop loading in it.

If a connection stops working

An app that signs in through your browser renews its sign-in on its own. If the same renewal is sent twice at the same moment, the second one is simply ignored and the agent keeps working. If an old, already-replaced renewal is sent again later, Gryz treats it as a possible copy and ends that agent's sign-ins, so you reconnect the app once. Owners and admins see this in the company Activity log.

Managing and stopping an agent

Owners and admins see every agent in the company and can pause, resume or remove any of them. A project manager can give an agent access to a project they manage, or take it away. Changes apply on the agent's next request, and every change is in the company Activity log.