AgnView

One screen for the coding agents already running on your machine.

What AgnView is

AgnView is a local dashboard and coordination hub for the agent CLIs you have installed: Claude Code, Codex, AntiGravity, DeepSeek, and any local model you run yourself. There are five known agent roles a task can be assigned to: claude_code, codex, antigravity, deepseek and custom. It does three things.

Watch. Every agent's output streams into one console, colour-coded by agent, instead of being scattered across terminal tabs.

Coordinate. Describe a job as a set of tasks with dependencies. A task stays blocked until the tasks it depends on complete, and when it starts, the agent picks it up with the upstream summaries and artefacts attached. If one agent finds a defect in another's work, it can send the task back for revision, or mark it failed, which blocks everything downstream.

Keep an eye on quotas. Session and weekly limits for your Claude, ChatGPT, Gemini, AntiGravity and DeepSeek accounts, in one place.

Every figure comes from the tool's own sign-in on this machine. AgnView reads the Claude Code, Codex and Gemini credentials, asks each provider for the real windows, and shows which source a figure came from and how long ago it was measured. A provider that publishes no figure says so, with the reason, rather than showing a number nobody measured. Nothing is ever written back to those credential files, so a signed-out or expired tool reports that instead of being refreshed behind your back.

AgnView runs entirely on your own machine. It is free, and there is nothing to sign up for.

Screenshots

	Light	Dark
Pipelines
Usage

Supported platforms

Clients

The AgnView hub in this repo covers Windows, macOS and Linux.

Mobile companion apps are in development. Links to their App Store and Google Play listings will be added here once they are published.

AgnView does not install any coding agent for you. It watches and dispatches to the CLIs already on your machine: Claude Code, Codex, AntiGravity or any other tool you add as an adapter. Install the CLIs you want to use yourself, before or after installing AgnView.

Install

The quickest way is not to install at all. uv fetches and runs AgnView in one command, with no clone and no virtualenv:

uvx agnview serve

To keep it around permanently:

uv tool install agnview      # or: pipx install agnview / pip install agnview

Working on AgnView itself? Install from source:

git clone https://github.com/tlaskar-git/AgnView.git
cd AgnView
pip install -e .

agent-relay is supported as an alias for the agnview command.

Windows desktop app

On Windows, AgnView also runs as a desktop app with its own window, so no browser is needed. Pick one of three ways to install it.

Download the app. No Python is needed. Download AgnView-windows-x64.zip from the latest release, extract the AgnView folder to %LOCALAPPDATA%\Programs\AgnView and run AgnView.exe. Keep the whole folder together, because the exe needs the files beside it.

Install from PyPI. This needs Python 3.10 or later:

pip install "agnview[desktop]"
agnview-desktop

Build it from source. From a clone of this repository:

.\tools\build-windows.ps1 -Install

This builds AgnView.exe, copies it to %LOCALAPPDATA%\Programs\AgnView and adds a Start menu shortcut.

The app needs the Microsoft Edge WebView2 runtime, which Windows 11 and current Windows Server include. It runs the hub on loopback and shows the dashboard in its window. Closing the window, from the title bar, the taskbar or with Alt+F4, quits AgnView and stops the hub.

The tray menu has a Start with Windows toggle. It is on by default. It creates a Task Scheduler sign-in task for your account, which starts AgnView hidden in the tray 15 seconds after you sign in. Open it from the tray icon or the Start menu. Settings live in ~/.agnview/desktop.json. Pass --port once to change the port, and the app remembers it.

Start

agnview serve

The dashboard is then at http://127.0.0.1:8765.

By default AgnView binds 127.0.0.1, so only your own machine can reach it. To let your phone or another computer on the same network connect, opt in explicitly:

agnview serve --listen-lan

Choose a different port with --port, and require a token from connecting clients with --token.

The hub starts automatically at login. The first agnview serve registers the OS-native autostart mechanism (the Windows Registry Run key, a macOS LaunchAgent, or a Linux XDG autostart entry, depending on your platform) with the options you served under, so a hub started with --listen-lan or a custom --port comes back the same way after a reboot. Change those options and the next agnview serve updates the registration. The pairing token is never written into it: the hub reads the persisted one at every start.

agnview autostart disable removes it and remembers that you did, so later starts leave it off. agnview autostart enable puts it back, and agnview autostart status shows whether it is currently registered. The switch in the dashboard settings does the same thing.

Pair a phone

1. Start the hub with agnview serve --listen-lan.
2. Open the dashboard and choose the mobile option in the header to show the pairing QR code.
3. Scan it with the AgnView app on a device on the same network.

The QR code carries the hub's LAN address, its certificate fingerprint, and a pairing key. Regenerating the code invalidates every device already paired. docs/PAIRING.md is the full contract.

Treat the pairing QR code like a password. Anyone who can read it can reach the hub, and the hub can run commands on this machine. See SECURITY.md.

LAN pairing, plus a remote path if you want it

On your own local network, your machine and your phone talk to each other directly. Nothing goes through anybody else's infrastructure, and no account or setup is needed.

If your phone is off your home or office Wi-Fi, LAN pairing alone cannot reach it. AgnView also accepts connections over iroh, which can carry the pairing QR code and the console beyond the LAN without a VPN. See "Remote access" below for what that depends on and how to opt out of it.

Remote access

docs/REMOTE-ACCESS.md covers every path in order: LAN, bundled iroh, what iroh depends on, running your own relay, and the overlay alternatives (Tailscale, NetBird, an SSH local port forward, or an existing WireGuard tunnel) via --listen-overlay.

Remote access uses iroh's free public relays. Most connections are direct and never touch a relay. Relayed connections are rate limited and carry no uptime guarantee. You can point AgnView at your own relay in Settings.

Docker, and what a container cannot do

docker compose up -d

The hub is then at http://localhost:8765. The image runs as an unprivileged agnview user and carries a HEALTHCHECK against the dashboard root, and the compose file keeps ~/.agnview (pairing token, adapters, notification config) in a named volume so it survives rebuilds and restarts.

A containerised hub cannot dispatch to CLI agents. AgnView dispatches work by executing the agent binaries on the host: claude, codex, antigravity and the rest, with their own credentials, config and your working tree. A container sees none of that.

• Docker suits watching and managing a pipeline: the dashboard, the REST and SSE API, job and dependency state, web-LLM prompt generation, quota monitoring, and mobile pairing, with remote worker daemons doing the actual execution.
• A native install suits dispatch. If you want AgnView to run Claude Code, Codex or AntiGravity for you, install it on the machine where those CLIs live.

Command line

agnview serve runs the hub. The rest of the commands drive it, so an agent can take part in a pipeline from a script:

Run any of them with --help for the full set of options.

How a pipeline runs

1. Create a job whose tasks name their dependencies. Tasks with no unmet dependencies start ready; the rest start pending.
2. An agent claims a ready task, which moves it to in progress.
3. Completing a task records its summary and artefacts, and unlocks every task whose dependencies are now met.
4. An agent that finds a defect upstream requests a revision. The upstream task reopens and its dependents go back to pending until it is fixed.
5. A task that cannot be finished is marked failed, which blocks everything downstream of it and fails the job.

Every one of these changes is pushed to open dashboards over Server-Sent Events, so a second tab or a paired phone follows along without a reload.

Licence

MIT, as declared in pyproject.toml. The repository does not yet carry a LICENSE file; one should be added before public release.