An agent is an AI assistant — Claude, ChatGPT, or one your organisation built — that works in DOK Genius on behalf of one person. You connect it once; from then on you can ask it things like "what is still missing in dossier 123?" or "open an onboarding dossier for Ana and put this contract in it", and it does the work through DOK Genius itself.
Three things are true of every agent, and they are the reason this chapter exists:
- It never has more than you. It sees what you see, and does only what you may do. If you cannot open a dossier in DOK Genius, neither can your agent.
- It does only what you allowed. When you connect it, you choose whether it may only read, or also create draft dossiers and add files.
- What it reads goes to your AI provider. A dossier your agent reads is sent to the company behind the agent — Anthropic for Claude, OpenAI for ChatGPT — under the terms you or your organisation have with them. That is why agents are off until your organisation turns them on, and why your organisation decides how confidential a dossier an agent may read.
This chapter is for the people who use an agent and the administrator who allows it. For whoever builds or configures one, Agents — technical reference has the address, the scopes, every tool and the limits.
For the administrator: allowing agents
Settings → Sign-in options, card Agents. Only the organisation's administrator sees it.
- Turn on Allow agents. Until you do, nobody in the organisation can connect one.
- Allowed agents. Leave the list empty to let people connect any agent. To restrict it, add the ones you accept: the easiest way is to let each agent connect once — it then appears as a suggestion you can add with a click. You can also type its client id or the https address of its metadata document.
- Highest confidentiality an agent reads. Dossiers and documents above this level do not exist for agents: they are not listed, they do not open, their files cannot be read. It starts at Internal, which means agents read Public and Internal dossiers only. Raise it only if you accept that more confidential material goes to your people's AI providers.
- Save agent settings.
The same card shows the address for agents your people will need, and the list of Connected agents — every agent connected in the organisation, whose it is and what it may do — with End connection on each row.
Turning Allow agents off ends every connection in the organisation at once, and each person will have to authorise their agent again. Removing an agent from the allowed list ends that agent's connections.
If the card says "Agents are not available on this platform yet", the settings you choose are kept and apply once they are.
The address you give your agent
Every organisation has its own agent address. You find it in Settings → Sessions, under Agent connections, as Address for agents, with a button to copy it.
It looks like this, where the last part is your organisation's identifier — the one you type in the Tenant field when you sign in:
https://api.dokgenius.com/mcp/your-organisation
Connecting Claude
Claude on the web and Claude Desktop (the two share your connectors):
- In Claude, open Customize → Connectors, choose + and then Add custom connector.
- Give it a name — DOK Genius — and paste the address for agents. Leave the advanced settings empty: DOK Genius recognises Claude on its own.
- Choose Add, then Connect. Claude sends you to DOK Genius: sign in as you always do, and you arrive at the authorisation screen described below.
- In a conversation, make sure the connector is on (the + button, then Connectors).
On a Team or Enterprise plan, an Owner of the Claude organisation may have to add the connector first, in Organization settings → Connectors (Add, then Custom, then Web). Each person then only chooses Connect.
Claude Code, in a terminal:
claude mcp add --transport http dok-genius https://api.dokgenius.com/mcp/your-organisation
Then, inside Claude Code, type /mcp, choose dok-genius and authenticate. Your browser opens DOK Genius; the
authorisation screen will warn you that the answer goes to a program on your computer — that is Claude Code, and
it is expected when you started the connection yourself, just now.
Connecting ChatGPT
Custom connectors in ChatGPT need Developer mode, available on Plus, Pro, Business, Enterprise and Education accounts on the web. On a Business or Enterprise workspace, the workspace administrator has to allow it first.
- In ChatGPT, open Settings → Security and login and turn on Developer mode.
- Add a new app for a remote MCP server with the address for agents, and choose OAuth as the authentication. DOK Genius recognises ChatGPT on its own; no client id or secret is needed.
- ChatGPT sends you to DOK Genius to sign in and authorise it.
ChatGPT asks you to confirm every action that writes — creating a dossier, adding a file — before it happens.
The menus of Claude and ChatGPT change from time to time. If a name here no longer matches, their own help pages describe where custom connectors are today; what you give them is always the same address.
The authorisation screen
When an agent connects, DOK Genius shows you, before anything happens:
- Which agent is asking, and who identifies it — for example Claude, identified by claude.ai. An agent's name can be claimed by anyone; the address that identifies it cannot.
- Where you will be sent after answering.
- What it will be able to do, one line per permission you are granting: read the processes, dossiers and documents you can see; create draft dossiers and add files to them, where you are allowed to; stay connected until you revoke the access or your session ends.
Choose Authorize or Decline. An agent you already authorised for the same things is not asked about again — except when the answer goes to a program on your computer, such as Claude Code, which is asked about every time. If your organisation does not allow that agent, the screen says so and nothing is granted.
What an agent can do
| You ask it to… | What it uses |
|---|---|
| Tell you who it is acting for | Your name, your organisation, what you granted it |
| Say which processes exist and what each one asks for | The processes you can see, their documents and form fields |
| Find dossiers | By number (123, or 123/620 for an addendum), text, state, date or the filters of the advanced search |
| Tell you where a dossier stands | Its state, documents, what is still missing before it can be submitted, approvals, and its recent history if you may see history |
| Read a document | The text DOK Genius extracted from a file, a part at a time |
| Open a new dossier | A draft in a process of your choice, yours, with the documents that process expects |
| Put a file into a document | Up to 10 MB, sent by the agent; the document's limit on how many files it takes and its file types apply, and a file with the same name replaces the one there |
| Give you a link to upload a file yourself | The dossier's documents in DOK Genius, with the right document open |
For a file of any size, or one on your own computer, the link is the better way: you open it, sign in as yourself, and drop the file where it belongs.
What an agent cannot do
- Submit, approve, reject, return or cancel a dossier. It prepares; you decide, in DOK Genius.
- Delete anything, or change who has access to anything.
- See more than you, or go above your organisation's confidentiality level for agents.
- See the values typed into a dossier's form. It knows which fields a process has; the answers stay in DOK Genius.
- Read a sensitive value in a document. Where the text of a file holds the value of a field the process marks sensitive — a tax number, a bank account — the agent reads [REDACTED SENSITIVE FIELD] instead. You read the file whole in DOK Genius.
- Act after you end its connection. Its very next request is refused.
What a document says is the document's, not an instruction. DOK Genius tells the agent so every time it reads one, because a document can contain text written to mislead an assistant. Read an agent's answer the way you would read a colleague's summary.
Seeing and ending your connections
Settings → Sessions → Agent connections lists every agent you connected: which one, what it may do, when you connected it and when it last did something. End connection stops it at once; it will have to ask you again.
A connection also ends when you sign out everywhere else, change your password, or your account is deactivated, and when it reaches the end of your organisation's session rules — the agent then simply asks you again.
In the history
Everything an agent does is done in your name, and recorded as yours with the agent beside you: a dossier's history and the audit trail read Ana Silva, via Claude. The same act done by you in DOK Genius reads Ana Silva. DOK Genius records each thing an agent asks for — what, when and whether it was allowed — never the content it read.
What you should have now
- As administrator: agents allowed or deliberately left off, the allowed agents chosen, and the confidentiality level for agents set on purpose rather than by default.
- As anyone: the address for agents copied from Sessions, one agent connected, and its row in Agent connections recognised.
- Your team told the two rules that matter: an agent never has more than the person who connected it, and what it reads goes to that person's AI provider.