The Assistant
The Assistant is a personal assistant that keeps its conversation with you
across sessions. It runs on your computer as a persistent agent: you start
it, talk to it, stop it, and when you start it again the conversation
continues.
This is the package that runs
The Assistant is built into the Agentry CLI and the desktop app, and it runs
from there: agentry agent start assistant, then agentry agent chat assistant, in the CLI. This release has the same bytes. The registry built
its archive from the package the CLI carries, and the archive's digest is the
one the CLI reports. Run agentry builtins to print it and compare it with
the digest on this page.
Installing this release does not give you the Assistant. A persistent release
runs only as a built-in, so the application runs it, with what is listed
below; the runner refuses to run this release as an ordinary QuickAgent.
What the application gives it
The manifest declares no network access, no file access, no commands and no
secret. When the application runs the Assistant, it gives it what follows,
and agentry agent show assistant lists what it has:
- Its conversation is kept on a volume of the agent's own, which the
runner keeps between starts and removes only when you reset the agent
(
agentry agent reset assistant).
- Its own files are on a second volume, its data folder, kept the same
way.
- Its memory is on a third volume, kept the same way: pages it writes in
memory/ for itself to read, on what you tell it and what you share with
it, one fact per line with its source (a file in your share and the date it
read it, or what you said and when), an index and a log. It reads its memory
at the start of each topic and adds to it when a topic ends. When its
conversation has grown large (two hours with nothing from you and 100,000
tokens or more, or three quarters of the model's context window in use), it
saves the conversation to its memory and goes on in a fresh conversation
that opens with a short handoff it wrote; agentry agent compact assistant
does it at once. Once a day, after two hours with nothing from you, it
tidies its memory if it changed (duplicates merged, long pages split, the
index kept short), as one of the same upkeep turns. Agentry reminds it to
read its memory at the first message
of each new conversation and when you come back after half an hour. Its
Agentry tool recall searches the conversations it had before, compacted
ones included, as the transcript on your computer keeps them (your messages,
its answers and its tasks' turns, never tool results); it changes nothing.
It is told never to change your share's files to keep it.
When the memory is made, the notes in its data folder's notes/ are copied
into it, and stay where they are. agentry agent memory export assistant <folder> copies the memory into a new folder, readable by you alone; a
reset removes it.
- Eighteen skills of its own, on unless you switch them off:
memory-ingest, memory-query, memory-consolidate and memory-compact,
which say how it adds to its memory, answers from it, tidies it and saves a
whole conversation before it is compacted; meeting-notes, which says how
it summarises a meeting Agentry Desktop recorded and transcribed (a call,
or a meeting held in the room, whose voices it never names unless the
transcript does), each point with its time in the transcript, saves the
summary beside it, files the
meeting in its memory, and arms the recordings you ask it for; google-workspace-setup, which
walks you through connecting your Google account; google-workspace,
google-inbox-triage and google-meeting-prep, adapted from gog's own
skills (MIT, their licence beside them), which say how it uses your Google
account safely; microsoft-365-setup, which walks you through connecting
your Microsoft account with an Entra app of your own; microsoft-365,
microsoft-inbox-triage and microsoft-meeting-prep, which say how it uses
your Microsoft account safely; fastmail-setup and
fastmail-inbox-triage, for your Fastmail account;
calendar-feeds-setup and calendar-feeds, which walk you through giving
it a calendar by its iCal address and say how it reads one; and
quickagent-use, which says how it works out what a QuickAgent needs
from what it shows (its questions, files, keys and guide) and runs it for
you.
- A share folder of yours,
~/agentry/assistantshare unless you choose
another, is mounted read and write while it runs: you see what it writes
there as it writes it. A file you add is checked for malware with ClamAV
before it reads it, and one ClamAV finds is not read; a file the check
cannot read (no signatures on your computer, or past its size limits) is
read unchecked, and the segment's receipts say so. You can switch the share
off.
- MCP servers it comes with, each an entry of its roster that you switch
on and off: Wikipedia (
@cyanheads/wikipedia-mcp-server, English
Wikipedia alone), on until you switch it off, and web search with Brave
(@brave/brave-search-mcp-server), off until you store your own Brave
Search API key for it (agentry agent secret set assistant BRAVE_API_KEY).
Each reaches only its provider's host. It also suggests two servers you can
add as your own, for your own use under their providers' terms: weather
from Open-Meteo, whose free service is for non-commercial use, and maps
from OpenStreetMap, whose public servers ask for light use
(agentry agent configure assistant --add-suggested weather or maps).
- Google Workspace through gog's MCP server, run from the image Agentry
pins, off until you switch it on and sign it in with your own Google Cloud
OAuth client (in Agentry Desktop, the Google card's Set up in its
Configuration's Mail and calendar; in a terminal,
agentry agent configure assistant --google-client <file>, then --google-sign-in <email>); the
google-workspace-setup skill walks you through it. It reads Gmail, Drive,
Docs, Sheets and your Calendar's events, read only unless you also switch
on its changes, and sends mail only with sending switched on too. It keeps
its sign-in on a volume of its own, under a password kept in your system
keychain, and reaches Google's hosts alone.
- Microsoft 365 and Outlook.com through Softeria's MCP server, run from
the image Agentry pins, off until you switch it on and sign it in with your
own Entra app (in Agentry Desktop, the Microsoft card's Set up in its
Configuration's Mail and calendar; in a terminal,
agentry agent configure assistant --microsoft-client <id>, then --microsoft-sign-in <email>), by
a device code. A work or school account usually needs its administrator to
approve the app first, and the microsoft-365-setup skill writes that
request with you. It reads your mail, calendar, contacts and OneDrive, read
only unless you also switch on its changes, and sends mail only with
sending switched on too. It keeps its sign-in on a volume of its own,
sealed under a key kept in your system keychain, and reaches Microsoft's
sign-in and Graph hosts alone.
- Fastmail through Fastmail's own MCP server, run at Fastmail and
reached at
api.fastmail.com alone, off until you give it your own
Fastmail API token of the MCP type (in Agentry Desktop, the Fastmail card's
Set up in its Configuration's Mail and calendar; in a terminal, agentry agent configure assistant --fastmail-token), which is kept in your system
keychain once Fastmail
accepts it: it can read, change or send in your account whatever your token
allows, so make the token Read only unless you want changes.
- Mail providers: Google, Microsoft and Fastmail are each enabled or disabled
(
agentry agent configure assistant --mail-provider <key>); only an
enabled one's server is on its roster, and setting one up enables it.
- Calendar feeds: any calendar you give it by its iCal (ICS) address,
under a name you choose (in Agentry Desktop 0.6.176 or later, Add a
calendar feed in its Configuration's Mail and calendar; in a terminal,
agentry agent configure assistant --calendar-feed <name>, the address
read without echo); the calendar-feeds-setup skill walks you through it.
It takes Google Calendar's secret address, the ICS link Outlook gives a
published calendar, an iCloud public calendar or another service's. An
address that answers only a signed-in browser, such as a Google Workspace
calendar's public address while its secret one is turned off, is given
with --calendar-sign-in: a browser window opens at it, you sign in, and
the cookies that window sends to the calendar's site (your sign-in to that
site, not a key to one calendar) are kept in your keychain beside the
address, sent to that site alone. Each feed is an entry of its roster, on once
you agree.
Its tools calendar_events, calendar_free_time and calendar_feeds read
them read only, in your time zone, fetching each from its own host alone
and at most every five minutes; it never changes your calendar, and
proposes an event as an .ics file in your share folder for you to open.
The address is kept in your system keychain and given to its tools server
alone: the agent is never shown it.
- What you add: knowledge packs from your computer and MCP servers, each
an entry of its roster that you switch on and off. Adding one, or turning
on one you have not agreed to on this computer, asks you first.
- A name and a personality, when you give them, reach it at its next
message. On its first conversation it asks you for both, and saves your
answer through its Agentry tool
self_configure, which changes its name
and personality and nothing else.
- Agentry Helper's tools, in Agentry Desktop: the same Agentry tools
Agentry Helper has, so the Assistant can do anything the Helper can.
registry_search and quickagent_show find a QuickAgent and say what it
may reach and how to run it (its questions, the files it takes, the keys
it needs and its publisher's guide), quickagent_install installs one,
quickagent_run runs an installed one with answers to its questions,
files from its share folder and a task, and run_result reads the
outcome; its quickagent-use skill says how it works out what to give a
QuickAgent from what you asked; installed_list lists what is installed, schedule_add,
schedule_list, schedule_pause and schedule_remove keep scheduled
runs, settings_open opens a section of Settings for you, and docs_read
reads a page of Agentry's documentation. The desktop asks you first, in a
dialog that names the Assistant, before an install, a QuickAgent's first
run for it, a run that fetches from an address you have not agreed to for
that QuickAgent, a schedule and a schedule's removal, and a No does
nothing.
None of these tools changes a setting, starts or stops an agent, signs in,
publishes or touches a key. Run from the CLI, these tools answer that only
Agentry Desktop answers them.
- Meeting recordings, in Agentry Desktop with its Meeting notes on: when
its roster's meetings entry is on (off until you switch it on, and only
with its heartbeat on), its Agentry tools
meeting_arm, meeting_list and
meeting_cancel arm Agentry Desktop to record a call during a meeting on
your calendar that you asked it to record. A window opens 5 minutes before
the meeting and closes 15 minutes after it, records only a call the desktop
detects inside it, and is refused when it would open more than an hour
ahead, so the Assistant arms each meeting from a task in the hour before it.
Each armed window shows in Agentry, where you can cancel it. Off, or run
from the CLI, these tools record nothing.
- The time: each message it is sent starts with the time now and your
time zone.
- Tasks and a heartbeat. When you ask it to do something at a time, it
registers a task through its Agentry tools (
task_add, task_list,
task_update, task_remove): a name, a cron time and an instruction to
itself. While it runs and its roster's heartbeat entry is on (off until you
switch it on), its heartbeat fires each task as a message of its own at its
time, which uses your model plan. A task due while it was stopped does not
run: it is told at its next message. At most 25 tasks, none more often
than every 15 minutes and 96 fires a day in all. agentry agent tasks assistant lists them, and pauses, resumes or removes one. A task gives it
nothing it does not already have.
- Each start is a segment, with its own receipts and result envelope in
the agent's folder under the runner home, and the conversation's record
continues across segments.
- It runs without the session caps an interactive QuickAgent has: no idle
timeout, no time limit and no turn limit. You stop it.
- It runs on the harness and model your runner policy names, in a container,
with the same sandbox every run has: no route out but the model's.
Using Agentry: https://docs.secureagentry.net/