AG2 Assistant

An open-source personal AI assistant powered by AG2's Beta framework, built in Python.

Public product name: AG2 Assistant. ag2-assistant remains the internal package, CLI command, and data-dir name.

AG2 Assistant is a web app (with optional messaging-channel and CLI front-ends) that acts as your personal AI agent — it searches the web, runs code, generates images, reads your project files, and runs scheduled tasks, all while showing its work in the open. Every reply is a projection of the underlying AG2 event stream.

Quick Start

1. Install

The quickest way — one line, no clone, no Python setup (the script installs uv if needed, and uv brings its own Python):

curl -fsSL https://raw.githubusercontent.com/ag2ai/ag2-assistant/main/scripts/install.sh | sh

On Windows:

powershell -ExecutionPolicy Bypass -c "irm https://raw.githubusercontent.com/ag2ai/ag2-assistant/main/scripts/install.ps1 | iex"

Re-run the same script any time to upgrade. Or clone for a development checkout:

git clone https://github.com/ag2ai/ag2-assistant.git
cd ag2-assistant
python -m venv .venv
source .venv/bin/activate
pip install -e ".[dev]"

The repo also ships a uv.lock, so uv run ag2-assistant … and uv run pytest work without setting up a venv yourself.

Or just install the CLI globally yourself, no clone (git = latest commit; releases publish to PyPI from the first tagged version):

uv tool install "git+https://github.com/ag2ai/ag2-assistant.git"   # or: pipx install "git+https://github.com/ag2ai/ag2-assistant.git"

Prefer to run it as an always-on container? See Deployment.

2. Run

ag2-assistant run        # gateway + web UI (+ any configured channels), one process

Then open http://localhost:8800/ and follow the first-run setup — six short steps:

Welcome — your name.
About you — location, working hours, preferred answer style (all optional, shared by every profile).
Connect — add a provider key (Gemini is the default and recommended — get one at aistudio.google.com), or sign in with ChatGPT to run on your existing ChatGPT subscription instead of an API key. You can't continue until one of the two works.
Profiles — create at least one profile (e.g. Personal), each a colour-coded, fully isolated workspace.
Set up — per profile: an optional folder to grant it read access to, and the kinds of work you want help with.
Ready — pick a theme and start.

Everything here is changeable later in Settings.

Prefer to set the key up front? Put GEMINI_API_KEY=your-key in a .env (see .env.example) before running — keys can come from the environment, and setup will pick them up.

To serve the API/UI without any messaging channels, use ag2-assistant gateway instead of run.

Profiles

The assistant is organised into profiles — separate, colour-coded workspaces you switch between from the sidebar (or ⌘1–⌘9). A profile has its own chats, tasks, memory, files, folder grants, MCP servers and permissions, so a Work profile never sees your Personal one.

Provider keys, the model configuration, and who you are are shared install-wide (the global ~/.ag2assistant/config.yaml); everything else is per profile, overlaid from that profile's ~/.ag2assistant/profiles/<id>/config.yaml. State lives in ~/.ag2assistant/profiles/<id>/ and generated files in ~/Documents/AG2 Assistant/<Profile>/. Override the install root with --data-dir (or AG2ASSISTANT_DATA_DIR).

The web app

The primary interface is the Svelte web UI served at / (→ /app). It includes:

Chat — multi-turn conversations with live, streamed agent events (tool calls, code runs, web searches) rendered inline. Attach images or documents and ask about them.
Rich answers — where structure beats prose, the agent renders a live surface instead: a weather panel, market board, news digest, agenda, inbox brief, or decision matrix.
Steer or stop a turn while it runs — send a message mid-turn and the agent folds it into the work it's already doing, or hit Stop to end the turn — keeping whatever it produced up to that point, so the conversation carries on with that context.
Tasks — standing jobs the agent runs on a schedule. A task is just a prompt + an optional model + a schedule (hourly/daily/weekly/weekdays, custom cron, one-shot, or manual), editable in place at any time — or created for you by the agent from chat. Every run is a real chat: open it to watch the agent work, steer it mid-run, stop it, or keep talking in it after it finishes. Recent run outcomes feed the next run, and tasks created from Telegram/Slack/Discord report their results back there.
It asks when it's unsure — mid-task the agent can put a question to you (with tappable options) and resume with your answer, in the web UI or on a connected channel.
Image generation — generated images are saved to the shared workspace and shown as clickable inline thumbnails.
Files — browse, preview, download and delete everything the assistant has saved.
Memory — the assistant passively learns your preferences; 👍/👎 feedback (with a reason) feeds a memory-aware learner that dedupes and prunes conflicting notes.
Folders — register folders outside the workspace and grant a profile (or a single chat) read or read + write access, so the assistant can work with your code/notes. Access is granted, never blanket — a first-touch prompt can mint a grant on the fly.
Permissions — you decide what commands it can run: approve shell commands once or always.
Voice — talk to the assistant (Gemini Live / OpenAI realtime) over a browser audio bridge.
Usage — tokens and estimated cost, per profile and across the install.
Settings — models, API keys, voice, MCP servers, Google, channels, folders, permissions, profiles, and re-run setup.

Models

Set up as many named model configurations as you like (Gemini, OpenAI, Anthropic, Ollama, or any OpenAI-compatible endpoint), test them, and switch the active one at any time in Settings → Models. Two OpenAI paths are supported: a normal API key, or Sign in with ChatGPT, which runs on your ChatGPT/Codex subscription instead of paying per token (unofficial — see ag2-assistant auth --help).

Anthropic and Ollama need one extra install. Gemini, all three OpenAI paths, and the Claude Code / Codex CLI logins work out of the box. The other two ship as configuration types but their provider libraries aren't bundled — Settings → Models says so on the row and on the template card, so you find out before configuring, not at Test time:

pip install "ag2-assistant[anthropic]"   # Anthropic
pip install "ag2-assistant[ollama]"      # Ollama (local models)

Installed with the one-line script or uv tool install? Add the extra to that environment instead:

uv tool install --with "ag2-assistant[ollama]" "git+https://github.com/ag2ai/ag2-assistant.git"

Running in Docker? Nothing to install — the image already carries both libraries.

Claude Code (subscription) as the main model

The Claude Code · CLI login configuration type runs the assistant's main agent on your own Claude Code CLI over ACP — no API key, auth is the CLI's on-disk login (your Claude subscription). Requirements:

npm install -g @agentclientprotocol/claude-agent-acp   # the ACP adapter, must be on PATH
claude --version                                        # Claude Code CLI installed and logged in

Add a "Claude Code" config in Settings → Models. The model list is read from the adapter itself (no model names are hardcoded here or in the UI), and leaving it on CLI default uses whatever your CLI is configured for. What you get:

All assistant tools work. The agent's full tool registry (weather, web search, memory, channels, skills, MCP servers, …) is served to Claude Code over a local in-process MCP gateway (127.0.0.1, secret URL path). Tool calls execute through the assistant's own event stream, so permissions, approval prompts, and streaming tool cards behave exactly as with any other provider.
Workspace confinement. The ACP session's filesystem access is scoped to the profile workspace directory.
Advanced options on the config are ACP constructor overrides (turn_timeout, allow_terminal, …), not API sampling params.

Limitations: in Docker the assistant reaches the host CLI through the ACP bridge (AG2ASSISTANT_ACP_BRIDGE — container-only plumbing; don't set it for host runs) and tool exposure is disabled there; per-token usage/cost isn't reported for ACP turns. Note Anthropic's terms restrict distributing products that ride a claude.ai login — a locally-run personal instance driving your own CLI is the same shape as Zed's adapter.

Codex (ChatGPT subscription) as the main model

The Codex · CLI login configuration type is the same integration for OpenAI's Codex CLI: the main agent runs over ACP on your Codex CLI login (~/.codex), no API key. Requirements:

npm install -g @agentclientprotocol/codex-acp   # the ACP adapter, must be on PATH
codex login status                               # Codex CLI logged in (ChatGPT subscription)

Add a "Codex" config in Settings → Models. Model and reasoning effort are two selects, filled from the adapter's own catalog (Codex reports them as one name[reasoning] id, which is what gets stored); CLI default uses the CLI's own model. Everything from the Claude Code section applies unchanged: the assistant's tools are exposed over the local MCP gateway, filesystem access is confined to the profile workspace, Advanced options are ACP constructor overrides, and the same Docker-bridge and usage-reporting limitations hold.

Google

Connect Google in Settings and the assistant can search and read your Gmail, draft and send mail (with your approval), read and create calendar events, and read Drive/Docs/Sheets.

CLI

The web app is the main experience, but the CLI is handy for quick one-shots and scripting:

ag2-assistant run                       # everything in one process (gateway + UI + channels)
ag2-assistant gateway                   # REST + WebSocket API + web UI only
ag2-assistant chat                      # interactive multi-turn chat in the terminal
ag2-assistant agent "message"           # single-shot prompt → reply
ag2-assistant onboard                   # first-run interview (name, location, hours, style)
ag2-assistant version                   # show version

ag2-assistant profiles list             # profiles: list / create
ag2-assistant profile show              # the active profile's memory; `profile clear` wipes it
ag2-assistant permissions list          # folder + command grants: allow / revoke / block / unblock
ag2-assistant auth login                # sign in with a ChatGPT subscription (also: logout, status)
ag2-assistant google login              # connect Gmail/Calendar/Drive (also: logout, status)

ag2-assistant telegram                  # run on Telegram   (needs TELEGRAM_BOT_TOKEN)
ag2-assistant discord                   # run on Discord    (needs DISCORD_BOT_TOKEN)
ag2-assistant slack                     # run on Slack      (needs SLACK_BOT_TOKEN + SLACK_APP_TOKEN)

Channel tokens can also be pasted into Settings → Channels instead of the environment.

Examples

ag2-assistant agent "What's the current AG2 version? Search the web."
ag2-assistant agent "Calculate the first 20 Fibonacci numbers"
ag2-assistant agent "Compare FastAPI vs Flask. Search the web for current benchmarks."

Deployment

The Quick Start above is the developer setup. For other ways to run it — see docs/deployment.md for the full guide:

Tier	For	How
Contributor	hacking on the code	`git clone` + `pip install -e ".[dev]"`
CLI user	running locally, no clone	`curl -fsSL https://raw.githubusercontent.com/ag2ai/ag2-assistant/main/scripts/install.sh \| sh`
Self-hosted	an always-on instance	`docker compose up -d`
PyPI	released versions	`uv tool install ag2-assistant` (from the first release)

Docker (self-hosted)

No clone needed — tagged releases publish a prebuilt image to GHCR (:latest moves on each release):

docker run -d --name ag2-assistant -p 8800:8800 \
  -e TZ=Australia/Sydney \
  -v ag2_data:/data -v ag2_workspace:/workspace \
  ghcr.io/ag2ai/ag2-assistant:latest

Set TZ to your own zone — scheduled tasks use the container's local time, and the image defaults to UTC, so without it "remind me at 6am" means 6am UTC. The startup banner prints the resolved local time.

Or from a checkout, with Compose:

cp .env.example .env         # add a provider key (e.g. GEMINI_API_KEY) — optional
docker compose up -d         # build + run; open http://localhost:8800/

State and the agent's file workspace persist in named volumes. Code execution runs in-container by default (no host Docker socket); see docs/deployment.md for the docker-out-of-docker option and channel setup.

The gateway has no built-in auth. Only expose port 8800 beyond localhost behind a reverse proxy that handles TLS and authentication.

Running Tests

pytest                  # the default suite (integration tests are excluded automatically)
pytest -m integration   # only the tests that call a real LLM / network (needs API keys)
pytest -m ""            # everything

Architecture

docs/architecture.md is the full system design — every service, agent, endpoint, event type, data flow, and on-disk store — with a companion diagram, docs/architecture.svg:

A text sketch of the same shape:

  Web UI (Svelte)   ·   Messaging channels   ·   CLI
         |  REST + WebSocket (event stream)        |
    +-----------+
    |  Gateway  |  (FastAPI: /api/p/{profile}/message, /api/p/{profile}/stream,
    |           |   /api/health, /app)
    +-----------+
         |  per-profile, per-chat isolation; events streamed to every client
    +-----------+
    |   Agent   |  (AG2 + Gemini / OpenAI / Anthropic / Ollama)
    |   Tools   |  (web search, web fetch, weather, markets, shell, code exec,
    |           |   image gen, files, folder read/write, tasks/scheduling, Google,
    |           |   skills, your own MCP servers)
    +-----------+
         |
    +-----------+
    |  Memory   |  (preferences learned passively + from 👍/👎 feedback → SQLite)
    +-----------+

Install-wide state lives under ~/.ag2assistant/ — the global config.yaml (model configs live in its llm_configs: section), secrets.json (provider keys), and the profile registry; each profile keeps its own config.yaml overlay, chats, tasks and memory in ~/.ag2assistant/profiles/<id>/, and its generated files in ~/Documents/AG2 Assistant/<Profile>/. The install root is overridable with --data-dir / AG2ASSISTANT_DATA_DIR.

Project Status

Documentation

Architecture — full system design: services, agents, endpoints, event model, data flow (diagram)
Usage Guide — CLI commands, configuration, channels
Deployment — install tiers, Docker/Compose, self-hosting

Contributing

Contributions are welcome — see CONTRIBUTING.md to get set up and AGENTS.md for the development guidelines. To report a security issue, follow .github/SECURITY.md (please don't open a public issue).

License

Apache License 2.0

Name		Name	Last commit message	Last commit date
Latest commit History 459 Commits
.agents/skills		.agents/skills
.claude/skills		.claude/skills
.github		.github
docs		docs
scripts		scripts
src/assistant		src/assistant
tests		tests
web		web
.dockerignore		.dockerignore
.env.example		.env.example
.gitattributes		.gitattributes
.gitignore		.gitignore
.nvmrc		.nvmrc
.pre-commit-config.yaml		.pre-commit-config.yaml
AGENTS.md		AGENTS.md
CONTEXT.md		CONTEXT.md
CONTRIBUTING.md		CONTRIBUTING.md
Dockerfile		Dockerfile
LICENSE		LICENSE
README.md		README.md
docker-compose.yml		docker-compose.yml
pyproject.toml		pyproject.toml
skills-lock.json		skills-lock.json
uv.lock		uv.lock

Navigation Menu

Uh oh!

Repository files navigation

AG2 Assistant

Quick Start

1. Install

2. Run

Profiles

The web app

Models

Claude Code (subscription) as the main model

Codex (ChatGPT subscription) as the main model

Google

CLI

Examples

Deployment

Docker (self-hosted)

Running Tests

Architecture

Project Status

Documentation

Contributing

License

About

Releases

Packages

Contributors

Languages

Uh oh!

Folders and files

Latest commit

History

Repository files navigation

AG2 Assistant

Quick Start

1. Install

2. Run

Profiles

The web app

Models

Claude Code (subscription) as the main model

Codex (ChatGPT subscription) as the main model

Google

CLI

Examples

Deployment

Docker (self-hosted)

Running Tests

Architecture

Project Status

Documentation

Contributing

License

About

Resources

Contributing

Security policy

Stars

Watchers

Forks

Releases

Packages

Contributors

Languages