What agents may do

Your people can connect the assistants they work with, like Claude or Claude Code. Agent access signs the agent in as its person, so it can find a topic a colleague worked through, add to it, or leave its person a note. As an administrator, you decide the most any of these agents may do here.

There are two ways in. Claude Code, Codex or a script gets a token, a secret that lets it act as its person. Claude or ChatGPT signs in from its own app. Either way the agent works as one person and sees what they see. Its changes are theirs.

Two choices, and the smaller one wins

You choose on Agent access which tools any agent may use here. Every tool is allowed, new ones too, until you choose otherwise. Each person chooses on their own Security page, for each agent they connect. An agent gets only what both allow, so a person can allow less than you do, never more.

Emargen leaves a disallowed tool off the agent’s list, so the agent never tries it. If it calls one anyway, Emargen refuses and says where its person can allow it. A change to either choice applies from the agent’s next call.

We left out a switch for assistants alone. A token can be pasted into any tool, so such a switch wouldn’t hold. Your choice covers tokens and sign-ins alike.

The tools

Agents reach Emargen only through these tools, never the rest of the app. They come in five areas, from topics and notes to building an initiative’s slices.

Each tool either reads or changes something. A person picks one of three choices by name, or picks tools one by one. Read only is every tool that reads. Topics and notes gives every topic and note tool except importing a page and closing a todo. It also lets agents read published requirements and initiatives. Everything is every tool.

Area	Tool	Reads or changes	In Topics and notes
Topics	Find topics	Reads	yes
Topics	Read a topic	Reads	yes
Topics	Read the templates	Reads	yes
Topics	Read the organisation’s fields	Reads	yes
Topics	Start a topic or subtopic	Changes	yes
Topics	Write into a topic	Changes	yes
Topics	Upload files for a topic	Changes	yes
Topics	Add an app to a topic	Changes	yes
Topics	Save a template	Changes	yes
Topics	Save a field	Changes	yes
Topics	Import a page as a topic	Changes	no
Your notes	Search your notes	Reads	yes
Your notes	Read a note	Reads	yes
Your notes	Look back over sessions	Reads	yes
Your notes	Leave you a note	Changes	yes
Your notes	Close a todo	Changes	no
Building	Read slice work	Reads	no
Building	Claim a task	Changes	no
Building	Add a task	Changes	no
Building	Say what blocks a task	Changes	no
Building	Ask the initiative’s people	Changes	no
Building	Propose a change	Changes	no
Building	Close a task	Changes	no
Published requirements	Search published requirements	Reads	yes
Published requirements	Read a published requirement	Reads	yes
Published requirements	Export a requirement bundle	Reads	yes
Initiatives	Read an initiative	Reads	yes
Initiatives	Start an initiative and shape it on its own	Changes	no
Initiatives	Message an initiative or steer its run	Changes	no

A choice by name covers new tools of its kind as we add them. A list picked one by one never gains a tool by itself. A new assistant starts on Topics and notes, and a new token on Everything. The tools describes each one.

Agent access has no name for four of these tools yet.

Topics and notes lets an assistant save templates and fields, within what its person may do. Change a template and the topics started from it afterwards use the new one. Change a field and you reach topics already using it. And reading is already copying. An agent allowed to read a published requirement gets all of it, so the bundle tool only changes how it arrives.

No tool publishes or approves an initiative. We kept those acts for people, in the app.

Choosing for a job

• Where assistants draft customer answers from topics, Topics and notes is enough. Someone who only wants drafts picks Read only.
• Where Claude Code or Codex builds an initiative’s slices, pick Everything. Its building tools claim and close tasks and ask the initiative’s people.
• If you want no changes from outside at all, untick every tool that changes something on Agent access. Every agent can then only read, whatever its person chose.

Kept to one organisation

Each organisation has its own address for agents, on each person’s Security page. An agent given one organisation’s address works only there, and a token or sign-in for another organisation is refused. A key from another organisation, such as another company’s topic key, is refused too. Someone may connect the same assistant to two organisations. It asks which one a new topic or note is for, unless their words do.

Names and dates, never calls

Agent access lists every agent working as one of your people. Each row shows whose agent it is, whether it uses a token or a sign-in, its name, and when it was last used. It shows no calls. Nobody sees what an agent reads. Whoever a topic is shared with sees what an agent writes there, under the agent’s name. Each person sees how often their assistants used each tool over the last 30 days. Emargen keeps each agent’s calls for 90 days, for its person.

Stopping an agent

• One agent: select it on Agent access and Revoke. It stops at once. Its person can make a new token or connect again, and Emargen doesn’t tell them.
• One tool, for every agent: untick it on Agent access.
• Every agent: untick every tool.
• Someone leaving: their agents stop with them, since an agent only ever works as its person.

With single sign-on

single sign-on covers your assistants too. Someone connecting Claude or ChatGPT signs in through your company’s sign-in first. While passwords still work, a password sign-in will do.

Next

How to

• Decide what agents may do
• Connect your own agentas your people do.

Guides

• Connect Claude Code or Codex

Page by page

• Agent accesswhere the tools are chosen.