---

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 and Gemini subscriptions, in one place.

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

Screenshots

	Light	Dark
Console
Pipelines
Usage

Supported platforms

Windows	10 and 11
macOS	12 Monterey and later, Intel and Apple silicon
Linux	any distribution with Python 3.10 or newer
Python	3.10 or newer
Browser	any current Chrome, Edge, Firefox or Safari

The companion mobile client is iOS and iPadOS.

Clients

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

• iOS and iPadOS: github.com/tlaskar-git/AgnView-iOS holds the design prototype and mobile spec for the iOS and iPadOS client.
• Android: an Android client will live at github.com/tlaskar-git/AgnView-Android (reference only, not yet created).

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.

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.

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:

Command	Purpose
create-job	Create a job from a YAML or JSON pipeline file
status	Show job and task status
wait	Block until a task's dependencies are satisfied
claim	Claim a task and begin work
complete	Mark a task complete with a summary
reject	Send a task back to its agent for revision
my-tasks	List tasks assigned to an agent
nodes	List connected agent instances
worker	Poll for assigned tasks as a daemon
export-prompt	Build a context prompt for a web LLM
usage	Show live subscription quotas

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.