AbstractAssistant
AbstractAssistant is a gateway-native desktop assistant for the AbstractFramework ecosystem: a
macOS menu-bar app with a compact palette, a hands-free voice conversation mode, and a small CLI.
The desktop client stays thin. AbstractGateway owns workflow discovery, durable runs, provider
connections, tool execution and policy, workspace policy, and multimodal defaults. The assistant
keeps only local state: your chats, window placement, downloads, and the overrides you choose for
this app.
Tray / Palette / CLI -> AbstractGateway -> AbstractRuntime -> AbstractCore -> Providers
What you get
- A menu-bar palette whose title tells you what the app is doing, with a session switcher fed by
the gateway (your sessions from every device, optionally every client's) and a live gateway
connection orb.
- An optional local AbstractAvatar portrait in the palette,
with design and motion controls in Appearance. It follows
listening, speech, tools, waits and errors; it stays off until you enable it.
- Automations: schedule a conversation to run on the gateway at a fixed UTC interval or when an
email arrives (with the result emailed to you if you want), read its runs
as a chat, switch it off or on (Active), run it now, edit or archive it, answer the runs that wait for you, discuss a
result, and get a tray notification only when a result is notable, a run fails or a run waits for
you. See docs/automations.md.
- What waits for you, everywhere: the app follows your event stream on the gateway, so a line above
the transcript and the menu-bar badge show the approvals, questions and new automation results
across every conversation, within seconds. Open takes you to the conversation that waits. On
one Mac you get one notification per item: the gateway's tray icon notifies first and the
Assistant stays silent for what it already showed.
- Live run activity in the transcript: the current step, elapsed time, recent tool calls with
their arguments and durations, and pause / resume / stop controls.
- Live replies: with Stream replies on (the default), the
answer appears while the model writes it, with its reasoning folded away under Thinking;
the finished answer replaces the live text.
- Tool approvals in a modeless sheet that shows the gateway's risk tier for each call and never
blocks the palette; "always allow on this Mac" and per-chat trust for low-risk tools.
- Hold the microphone button to talk: release sends the words. A voice conversation mode (⌘⇧V,
or the microphone's right-click menu): listen, send, speak, listen again — with a live status
strip. Spoken replies stream from the gateway.
- A Settings window with seven sections: Connection, Models & reasoning (workflow, reasoning
effort, reply streaming), Voice, Workspace (my default workspaces, this chat), Tools & permissions,
Appearance & window, About. Every value says whether it is the gateway default or this app's
override. About is also on the menu-bar icon (About AbstractAssistant…).
- Local, persistent overrides for every route the assistant drives (chat, voice, image, video,
music, sound) that ride each request without touching the gateway's shared defaults.
- Per-answer statistics (tokens, tools, files, duration, model) with clickable detail views.
- One reconstruction of every run: on AbstractGateway 0.17.0 or later the desktop app,
assistant run
and the web Assistant (web/) render a run from the gateway's replay view (GET /runs/{id}/view
and its stream): status, replies, tool rows and the approval or question wherever it waits come
from the gateway, so an approval answered on another client closes here too. On an older gateway
the app folds the run's ledger itself and says so.
- Multi-attachment composer with an image gallery, drag and drop, and artifact previews.
- In your pocket and on your wrist: AbstractAssistant for iPhone and Apple Watch is one app
(0.16.0, pre-release through TestFlight): the iPhone app carries the watch app, each signs in
on its own and works without the other, and notifications reach you while the app is closed (Apple push, once your gateway
has a push forwarder). The iPhone has the same conversations, tools and automations, with tab badges for
what needs you (phone/iphone/README.md); the watch talks to the same
gateway by voice around the orb, answers approvals and questions, continues your sessions and runs
your automations, and your avatar on the watch face shows what needs you
(watch/apple/README.md).
Install
pip install "abstractassistant[voice]"
The voice extra adds local microphone capture (dictation and voice conversations). The base
install covers text chat, spoken replies and gateway-backed media; STT and TTS run on the gateway
either way.
Requirements: Python 3.10+, an AbstractGateway you can reach (0.16.0 or later for the voice
library and the chat-wide tool policy; an older gateway keeps the labeled local grant and refuses
the voice library's requests). AbstractCore is installed as a
dependency (its minimum version is set in pyproject.toml). macOS is the primary tray target;
Linux and Windows may work but are not packaged to the same standard.
Quick start
Start a local gateway for development:
export ABSTRACTGATEWAY_FLOWS_DIR="$PWD/abstractgateway/flows/bundles"
export ABSTRACTGATEWAY_AUTH_TOKEN="your-shared-token"
abstractgateway serve --host 127.0.0.1 --port 8080
Launch the tray assistant:
assistant
Run a single terminal turn:
assistant run --prompt "Search the web for the latest OpenAI news and summarize it with sources."
Connection overrides:
assistant --gateway-url http://127.0.0.1:8080 --gateway-token "$ABSTRACTGATEWAY_AUTH_TOKEN"
--gateway-url on its own is enough when you have already signed in to that gateway: the app
reuses the sign-in saved in ~/.abstractassistant/gateway_connection.json.
If the gateway runs the AbstractGateway console, the simplest way to start the assistant is
Open on its card there: the console launches the app already signed in as you, and the app
remembers that sign-in for later launches. The console can sign in only an Assistant it launches,
so quit a running Assistant before you click Open.
Each turn runs the gateway's default workflow for the assistant (abstractassistant.agent.v1).
When the gateway sets none, the app runs its built-in orchestrator, the
abstractassistant-orchestrator workflow it publishes to your tenant catalog. You can pick
another assistant workflow in Settings → Workflow; the list holds only the assistant
workflows the gateway says you can run (what your admin made available, plus your own:
GET /api/gateway/bundles?executable_for=abstractassistant.agent.v1).
Defaults and durability
The gateway is the source of truth for the published workflow, provider and model defaults, tool
inventory and approval defaults, workspace policy, run history, waits and artifacts.
The desktop client stores under ~/.abstractassistant/:
preferences.json — this app's overrides and preferences (model routes, reasoning effort,
reply streaming, workspace grant, voice options, tool modes, hotkey, window size)
gateway_connection.json — gateway URL and sign-in state
session_cache.json, sessions/ — a rebuildable cache of the gateway's sessions (local labels,
cached transcripts, each session's workspace); the sessions themselves live on
the gateway
automations_notified.json — which automation notifications were already shown; automations
themselves live on the gateway
- downloads and cached audio
Documentation
Start with docs/README.md.
Development
pip install -e ".[dev]"
QT_QPA_PLATFORM=offscreen python -m pytest tests/basic -q
Project links