Skip to content

Agents and add-ins

An agent is the assistant inside another application: the side panel of the Fadenstack browser agent for Chrome, or the task pane of the Word, Excel, PowerPoint or Outlook add-in. The application brings the tools (read the page, fill a form, write a cell); Fadenstack decides who may use the agent, which model answers and which of the application's tools it may run. You set up one agent per application and grant it to the teams that should have it.

What people do with the agents is described in The browser agent and Office add-ins.

Agents → Agents: each agent with its address, description, status (Published or Retired), live version and when it was updated, and New agent at the top

Before you start

  • The application is installed for your users (see Install the applications).
  • A model the agent answers with: one of your own deployments, or a remote provider's. Everything a person asks, and the page or document the agent reads for it, goes to that model, so choose it with the same care as for the chat (see What leaves the organisation).
  • A group with the people in it (Access Control → Groups), or the people themselves: an agent is granted to people and groups, and refused to everyone else.

Create the agent

New agent with Chrome picked under Start from an app: the name, the address browser-assistant, the Model still to pick, Who may use it, and Publish it now ticked

  1. Open Agents → Agents and select New agent.
  2. Under Start from an app, pick the application: Chrome, Microsoft Word, Microsoft Excel, Microsoft PowerPoint or Microsoft Outlook. Its definition is filled in: the instructions, the tools the application brings with their classes, and the greeting.
  3. Keep the Address. It is what the application asks for, and it cannot be changed later:

    Application Address
    Chrome browser-assistant
    Word word-assistant
    Excel excel-assistant
    PowerPoint powerpoint-assistant
    Outlook outlook-assistant
  4. Pick the Model. It is required: the applications leave the model to the agent, and without one the agent cannot answer.

  5. Under Who may use it, type part of a group's name or an e-mail address, and pick it.
  6. Tick Publish it now to serve the agent to those people at once. Left unticked, it starts as a draft that only its maintainers can use.
  7. Select Create.

The agent's page opens. Its list status is Draft, Published or Retired.

The agent's page

The Word assistant's page on the Definition tab: the tabs Definition, Versions, Apps, Access, Skills and Activity, Edit and Retire at the top, the instructions and the model

Tab What it holds
Definition The version people get: Instructions, Model, Knowledge (collections it may search), Server tools (What a chat gets, Only these, or None), Local MCP servers (with Advanced → Allow plain http on the local network), Local skills, Host tools (the application's tools and their classes), Approval levels (the tool mode conversations start in, and the highest), Appearance (greeting and suggested prompts) and Privacy.
Versions Every version. Publish one, Roll back to an earlier one, see the Changes.
Access Who may use it, the Excluded people Enterprise and its Maintainers (with Enterprise: Maintainers and auditors Enterprise). Check someone's access says whether a person may use the agent, and why.
Skills The organisation's skills that come into the agent's conversations without anyone picking them: those assigned to everyone, and those assigned to this agent. For those who manage the agent.
Activity Usage over a time range (requests, errors, tokens, users, conversations: counts only, never content), shown only when at least five people used the agent in the range; the owner of a personal agent sees its figures all the same. What the applications reported (tools that ran, approvals, plain http switched on or off for a server), and the Signed-in devices, each with what its application last reported about it (see What a device reports). It names nobody but you.
Apps What the applications bring, compared with the agent's version (see When an application brings new tools).
More about server tools

Only these takes names or patterns (mcp.github.*, knowledge_search, current_time, get_available_tools); None lists nothing. The limit covers every tool the gateway runs for the agent: MCP tools, knowledge search, files attached to the conversation and the system tools. A tool outside it is not offered, is not found by the tool search, and is refused with an error if the model (or text the application passes in) names it anyway. The application's own tools (Host tools) and the files of a skill the agent uses are not affected.

More about local MCP servers

Local MCP servers: Not allowed (the default) keeps the agent to the server's tools and the application tools its version lists under Host tools; a version that lists none holds nothing back, but then every application tool is unclassified and asks first, every time. Allowed lets it attach MCP servers on the person's computer; their tools are unclassified, so they always ask first. People in Developer mode (see Users, groups and roles) may attach them either way.

Advanced → Allow plain http on the local network is off by default. Without it, the application reaches a person's MCP servers over https, or on their own computer. With it, the application may also reach a server over plain http at an address on the local network, once the person switches that on for the server and confirms a warning. Anyone on the network can read and change what goes to and from such a server: the tool calls, their arguments and their results, which may hold document content. Never use it with secrets: over plain http, the application refuses a server that needs a key or a sign-in. Developer mode does not override this setting. When a person switches plain http on or off for a server, the application reports it, with the server's address, on the Activity tab.

More about local skills

Local skills: Not allowed (the default) keeps the agent to your organisation's skills. The skills a person keeps on their own device stay off. Allowed lets the application use those skills too, beside your organisation's. They are marked as not reviewed, because nobody in your organisation has checked them; scripts in them never run. People in Developer mode (see Users, groups and roles) may use them either way. The setting only tells the application what it may do. The skills are kept on the person's device, not on the server; when the model uses one, its text becomes part of the conversation, like any other message.

More about an agent's skills

The Skills tab: For everyone in your organisation with one skill, and For this agent with two skills and the search to add more

For everyone in your organisation lists the skills assigned to everyone; they reach this agent too, and you change that on the skill's page. For this agent holds the skills assigned to this agent: find them by name or description among the published skills you administer, then Save. A skill you do not administer shows without a remove button: someone who manages that skill assigned it. The tab is not part of a version: a change applies at once, to every version. It is about your organisation's skills; Local skills on the Definition tab is about the skills people keep on their own device. Other skills still come in by their triggers, or when a person picks them.

More about chat sessions running at once

Chat sessions running at once (under Appearance) sets how many chat sessions of this agent can work at the same time on one device: choose 1 to 5, or Default (5). More sessions cost more and need more attention. The desktop agent follows this limit; the setting needs no other change.

The Versions tab of the Word assistant: v1, live, with its dates and View, below the agent's tabs and the Edit and Retire buttons

Saving the definition makes a new version (Save as new version, or Save and publish); versions never change. Publishing a version reaches everyone who uses the agent at their next request.

The Access tab: Who may use it, with Add a person or team and the person the agent is granted to, then its Maintainers and Check someone's access

Maintainers change the agent, publish versions and decide who may use it, and can use it before it is published. They do this from Agents in the chat's user menu; they need no access to the console.

Enterprise With Enterprise, an agent can also have Auditors: choose Add as → Auditor before picking the person, or change a member's role in the list, and select Save. Auditors see the agent's versions, members, settings and activity and change nothing; to use it, they need it granted like anyone else. Without a valid Enterprise licence they no longer see it.

To check someone's access, an auditor types the person's whole e-mail address: a part of one finds nobody, so the check does not list everyone (GDPR Art. 5(1)©, 25(2)). Every check is recorded in the audit log, with who checked whom (agent.access.check).

Works councils

Auditors read what people did with an agent. In Germany, a tool that can monitor employees' conduct or performance needs the works council's agreement (BetrVG § 87(1) No. 6): agree the auditor role with it before you give it.

Enterprise With Enterprise, Excluded people may not use the agent, even through a team that may: pick the person, give a Reason and select Exclude. It takes effect in the applications at once and is recorded in the audit log. The person is told only that they are excluded; the reason is for the agent's maintainers and administrators. Maintainers and administrators always keep the agent and cannot be excluded.

Without a valid Enterprise licence, an agent with excluded people can be used by its maintainers only, so that nobody excluded gets it back; the Access tab says so. Open to everyone granted lifts that: everyone the agent is granted to can use it again. Once the licence is valid again, the excluded people still listed are kept out again.

The Activity tab: requests, errors, tokens, users and conversations of the last 30 days with a chart, and below it what the applications reported, such as tools that ran and approvals

What a device reports

Each application reports, when it starts, the device it runs on: the operating system, the application's version, and the plugins and local MCP server programs it loads, each with the SHA-256 of its file. The Activity tab shows this under each signed-in device; each person sees their own devices under My profile. To find a changed program, compare its SHA-256 with the one your IT approved.

A program is kept by its file name, without its folder, so no user name comes in with a path. A device's report goes when the person's account is deleted, and 90 days after the device last reported.

When the application approves or runs a tool, it can also send the SHA-256 of the call's arguments. That value is kept with the event and with the decision record, never the arguments themselves. It matches the entry in the audit log the application keeps on the device.

Install the applications

Chrome. Install the extension for your users with Chrome's ExtensionInstallForcelist policy, and preset the server and the agent in the extension's managed storage (Chrome's 3rdparty policy), so that people only sign in and cannot change them:

{ "server": "https://ai.example.internal", "agent": "browser-assistant" }

server is the current name of the key; the older serverUrl is still read.

Add "language": "de" to fix the panel's language for everyone; without it, each person picks one in the panel's settings, and Automatic is the browser's language.

Add "webSearch": true to switch web search on for everyone, or false to keep it off; without it, each person decides in the panel's settings, and it is off until they do. Searches go to the search engine set in Chrome, not through Fadenstack; Chrome's DefaultSearchProviderSearchURL policy can point it at your own search service.

To keep the agent off some sites (online banking, HR systems), block them for the extension with Chrome's ExtensionSettings policy (runtime_blocked_hosts under the extension's ID): Chrome then lets it neither read nor act on those sites, whatever a person allows.

An agent made from the Chrome template gets the knowledge search and the time as its Server tools; add others, such as web search, in its Definition. Screenshots and pictures go to it only when its Model sees pictures.

Chrome uses the operating system's certificate store: the computers must trust the server's certificate (see HTTPS and trust).

Office. The add-ins run in desktop Office for Windows: Word, Excel, PowerPoint and Outlook, tested with the Microsoft 365 apps only. For Outlook, only classic Outlook: the new Outlook for Windows runs no such add-ins. IT presets the server and the agents in the machine policy %PROGRAMDATA%\Fadenstack\Office\policy.json; what it sets, people cannot change:

{
  "serverUrl": "https://ai.example.internal",
  "agents": {
    "word": "word-assistant",
    "excel": "excel-assistant",
    "powerpoint": "powerpoint-assistant",
    "outlook": "outlook-assistant"
  },
  "trustedCaCertificate": null
}

Add "language": "de" to fix the panes' language for everyone; without it, each person picks one in the pane's settings, and Automatic is Office's display language. The add-ins bring English, German, French, Spanish and Simplified Chinese: deploy the de, fr, es and zh-Hans folders beside the add-in, or the panes stay in English.

One sign-in serves all the Office add-ins on a computer.

More about trustedCaCertificate

trustedCaCertificate names a CA certificate file to trust for the server, for an installation whose CA is not deployed to the computers.

People sign in with their Fadenstack account. Their conversations stay on their computer; the server keeps usage and events (which tools ran, how long, with what outcome), never the content. What the agent reads reaches the model through Fadenstack, so your personal data policy applies.

Tools and approvals

Each tool an application brings has a class, which the agent's version sets:

Class In the tool modes
Read Runs in every tool mode but Off.
Write Asks first in Ask mode and runs in Auto.
Destructive Always asks.
Unclassified Always asks.

People pick the tool mode in the panel, per conversation. Under Approval levels the version says where every conversation starts (Starts in, Ask unless you change it) and the highest mode people may pick (At most, Auto unless you change it); the start is never above the highest, and your organisation's limits narrow both. Bypass can be the highest level only while an administrator allows it in Settings → Approvals, and a version that has it can be published only then; it can never be where conversations start. See Bypass. A research agent, for example, might start in Read only and go at most to Ask. The application can make a class stricter, never looser. Your tool rules apply to agents as well as chats.

Under Host tools, the version can also say how each tool asks (Approval):

Approval What it does
As the tool mode says The table above decides.
Always ask Asks every time; nothing the person allowed before, for the chat or always, answers it. Use it for tools that decide about people (GDPR Art. 22) or act on the physical world.
Ask Asks even where the tool mode would run the tool; the person may allow it for the chat or always.
Never ask A read or write tool runs without asking in Ask and Auto. A version where a destructive or unclassified tool has it is not published.

None of them shows a tool the tool mode hides. Until Fadenstack can mark tools that act on the physical world, such as a machine or a door, give them Always ask, or set an "Ask first" tool rule. Two examples: the Outlook add-in never sends mail itself, it opens drafts for the person; and the browser agent never reads or fills password fields, nor card fields the site marks as such.

Under Place a tool can say which of its arguments names where a call acts (a JSON pointer such as /path) and what kind of place it is: document, folder, website, app or desk. A grant for a stretch of work, which a person can give an agent for a while, covers the tool's calls only in the places it names. A tool without a place keeps asking.

When an application brings new tools

Each time an application starts, it tells the server what it brings: its tools with the class its developer gave them, the context it offers, its greeting and its version. The Apps tab compares this with the agent's version.

When an update of the application brings a tool the agent's version does not list, nothing breaks: the tool is held back (the model does not see it), and the application keeps working with the tools it may use. The dashboard shows "Tool changes waiting for …" until you decide.

An agent's Apps tab with nothing to decide: Version 2 has everything the apps bring, and each tool with its status In the contract, its class and the add-in that brings it

  1. Open the agent's Apps tab. Each tool shows a status:

    Status Meaning
    New The agent's version does not list it: it is held back.
    Changed Its parameters differ from the ones accepted.
    Riskier The application now calls it riskier than the agent does.
    Unused In the agent's version, but no application in use brings it.
    In the contract Nothing to do.
  2. Read the tool's description, especially when it carries a warning: Instructs the model, Hidden characters, Link or Long. Models read tool descriptions like instructions.

  3. Keep or change the class the application suggests, or choose Leave out.
  4. Select Accept and publish, or Accept as draft to publish later from Versions.

Nothing is taken in without that click. Apps in use lists each application and version that runs the agent, who last reported it and the context it offers.

Agents created from the extension

Someone whose role has agents:create can also start from the Chrome extension: when it finds no agent under its address, it offers Create it from this extension. The agent is created from the extension's definition as a draft they maintain. Publishing it and granting it to others is then done in the console or on their Agents page.

Turn an agent off

Turn off, at the top of the agent's page, stops its applications at once, for everyone, its maintainers and administrators too. Every chat and tool call is refused, and the applications say "Your organisation turned this agent off". Use it when something is wrong: an application misbehaves, a tool does what it should not, or the model must not be asked for a while. Those who manage the agent can turn it off, as they can retire it.

  • Grants end. The grants people gave the agent for a stretch of work end, and none of them counts again when the agent is back.
  • Sign-ins stay. Turn on brings the agent back as it was, and nobody has to sign in again.
  • The record stays complete. The applications still report what they did and the devices they run on.
  • It is said and recorded. The agent's page and list show Off, with when and by whom it was turned off. Check someone's access says the agent is off, and what would apply otherwise. Each switch is recorded in the audit log (agent.turn_off, agent.turn_on).

Retire also closes the agent, but its sign-ins stop working, and it comes before deleting the agent. Turn the agent off when it should come back soon.

If Turn on fails

Before the agent is back, the grants from before it was turned off are ended once more. When that cannot be done right now, the agent stays off and the page says so: try Turn on again in a moment.

Retire and delete

To Do this
Stop its applications for a while Turn off (see Turn an agent off); Turn on brings it back.
Close the agent to everyone at once Retire: sign-ins stop working, and its versions, access and history are kept. Restore brings it back as it was.
Delete a retired agent for good Delete, then type its address to confirm. Its versions, access and sign-ins go; what it did stays in the usage and audit logs, and its address is free for a new agent. Deleting needs the permission agents:delete, which the built-in admin role has and maintainers do not.
End one device's sign-in Sign out on the Activity tab: that person must sign in again.
How long agent sign-ins last

An agent sign-in on a device lasts 30 days, and ends after 7 days without use. Removing someone's access, deactivating them or turning the agent off takes effect at their next request. Signing out one device takes effect within five minutes: an application's access token lasts at most 300 seconds, and the application renews it with its sign-in. These are Agent sign-in lifetime (s), Agent sign-in idle timeout (s) and Agent access token lifetime (s) (60 to 300, 300 by default) under Settings → General.

Enterprise With Enterprise, an agent can have shorter ones: Sign-in lifetimes on its Definition tab, in seconds, at least 60; an empty field keeps the tenant's. They come with the version you save and publish. Longer ones than the tenant's are refused, and lowering the tenant's setting later shortens the agent's too. Devices already signed in get the shorter lifetime when their sign-in is next renewed. An access token lasts at most 300 seconds whatever is set.

When it does not work

The panel says why a sign-in failed:

You see What to do
There is no agent under that address. The address in the application differs from the agent's. Compare it with the Agents list; an agent made without Start from an app has an address derived from its name.
The person has no access. Add them, or their group, on the Access tab, and check that the agent is published. Check someone's access says what stops them, such as an exclusion.
Your organisation turned this agent off. Someone turned the agent off. Turn on on its page brings it back (see Turn an agent off).
The server does not answer. Check the server address in the application and that the computer trusts the server's certificate.
A tool the application offers is not used It is held back: accept it on the Apps tab.