Wardian

Wardian

Where local agent work becomes visible, durable, and malleable.

Run the CLI agents you already use in real workspaces. Keep their sessions and evidence visible, then turn successful work into reusable prompts, skills, classes, workflows, and project context.

Download Wardian GitHub Repository
Documentation Latest release Issues and support Discord
The Grid: every agent running in its own PTY-backed terminal, with status color on each card.

Wardian collects nothing

There is no Wardian server. Nothing about your work is sent anywhere, because there is nowhere for it to go.

Your work stays on your machine

Agents run as local processes in your own project folders. Prompts, skills, classes, workflows and results are files on your disk. The Dashboard reads a local database. Nothing is uploaded.

No account, no identifiers

There is no sign-up, no license check and no device id. This site sets no cookies and loads no analytics.

What does touch the network

Released builds ask GitHub whether a newer version exists. Otherwise it is only whatever your provider CLI connects to, on your own account, the same as it would without Wardian. Wardian never proxies that request, so if the CLI is pointed at a local model, nothing leaves the machine at all.

Open source is not the same as private. Plenty of open tools ship analytics. Wardian is MIT licensed and collects nothing, and the source is there to check. Read the code.

01 Graph

Agents that know about each other

The Graph draws the agents and the links between them. Edges come from a team you create, a link you make by hand, or a shared workspace, and you can ignore the ones you do not want.

The topology is written to topology.json, so an agent can read its own neighbors before it sends anything.

Graph guide
Graph view: topology, edges, neighbors, and team-seeded links.

02 Inbox

One place for whatever needs you

Finished work, automation outcomes, important updates, and the prompts an agent is blocked on all arrive in one list. Unread first, filterable by type, with a badge when something is waiting on a decision.

An approval request names the action, why it is risky, and the choices you can send back. It expires rather than proceeding on its own, and only one can be open per agent at a time.

Inbox guide
The Inbox: an approval request with its choices, and a completed result.

03 Two interfaces

A GUI for you, a CLI for your agents

You get a desktop app. Your agents get wardian, the same control surface as a command line: list peers, send a task, wait for the reply, read the result.

An agent can spawn a peer, ask it a question and block on the answer without a person relaying messages between two terminals.

Wardian CLI guide
an agent, using wardian 
$ wardian agent list
NAME            CLASS     STATUS
Docs-Codex      Coder     processing
Docs-Reviewer   Reviewer  idle
$ wardian ask Docs-Reviewer "does the capture manifest look right?"
waiting for reply...
reply    No blocking findings. Sizes are within budget.
$ wardian agent spawn --class Coder --name Docs-Encoder
spawned  Docs-Encoder  class=Coder
$ wardian memory save "Clips render near 1200px, so capture at 1280" \
    --evidence "a 1600px capture was shown at 0.4x and text turned to mush"
saved    id=7c1f9ab2  scope=workspace  revision=1

04 Workflows

Workflows that branch, loop, and wait

Build a run as a node graph: agent calls, commands, scripts, files, memory, branches, loops, and waits. The workflow owns the control flow, so a branch is decided by the run itself.

Input parameters and shared storage pass values between nodes, and the builder validates the graph before it executes.

Automations guide
The visual builder: control-flow nodes, parameters, and validation.

05 Dashboard

The fleet, not the tab

One row per agent: state, active time, turns, tokens, files touched, and lines changed, with a provider strip across the top.

Pick the window and the trend is computed over it.

Dashboard guide
The Dashboard fleet table and the Analytics trend metrics.

06 Garden

A map where position means something

Start with the shape of your work: agents, workspaces, and situated automations grouped into districts by the teams and projects they share. Zoom toward an agent to see its Identity, Skills, Memory, Automations, and Conversations in place.

Keep zooming to read an individual record and its evidence. Zoom back out to return to the agent and the wider habitat. The contents stay anchored as detail unfolds, so you can follow the path from the overview to the work behind it.

Garden guide
From the habitat to an agent's regions, into a memory record, and back out through the same map.

07 Files on disk

Everything is a file you can read

Prompts, classes, skills, and workflows are plain files in your Wardian home. So is the agent graph, in topology.json, and the queue behind the Inbox, in queue/items.json. The Library edits them and the Explorer browses the same directory.

The app is a view onto those files rather than a store that owns them. Edit topology.json in any editor and the Graph redraws while you watch. Close Wardian and the state is still there to read.

Explorer guide
A prompt in the Library, then the same file in the Explorer.

08 Memory

Memory that carries its evidence

Every saved memory stores the excerpt it came from. Records are scoped to an agent or a workspace, and an update keeps the previous revision rather than overwriting it.

Agent memory is opt-in under Settings, and inside a managed terminal an agent can only reach its own records.

Agent memory guide
wardian memory 
$ wardian memory save "Release tags are cut from main after CI is green" \
    --evidence "we tag from main, never from a feature branch" --scope agent
saved  id=7c1f9ab2  scope=agent  revision=1
$ wardian memory recall
ID        SCOPE      MEMORY
7c1f9ab2  agent      Release tags are cut from main after CI is green
a0d34e11  workspace  Run the verify script before every push
$ wardian memory show 7c1f9ab2
memory    Release tags are cut from main after CI is green
evidence  "we tag from main, never from a feature branch"
scope     agent      revision  1
$ wardian memory update 7c1f9ab2 "Tags are cut from main after CI is green and the changelog is written" \
    --evidence "add the changelog step before tagging"
updated  id=7c1f9ab2  revision=2
$ wardian memory history 7c1f9ab2
revision 2  current     add the changelog step before tagging
revision 1  superseded  we tag from main, never from a feature branch

09 Classes

Roles you define once

A Class is a role: instructions, skills, and prompts that every agent of that role receives when it starts.

The definition is a Library entry on disk, so editing the class changes the next agent that uses it.

Class management guide
Class management: a reusable agent setup being applied.

10 Remote

Check in from your phone

Wardian serves a mobile surface over your own private network. Read an agent's terminal, answer a prompt it is blocked on, and clear the Inbox from wherever you are.

The desktop stays the host: it keeps the agents, the PTYs, the filesystem access and the telemetry. The phone is a view onto it, reached over your own Tailscale tailnet with no Wardian server in between.

Remote control guide
Wardian on a phone showing a running agent's terminal with pause, clear and kill controls and a prompt composer.
An agent's live terminal, with its composer.
Wardian on a phone showing the Inbox with a completed agent result summary.
Completions in the Inbox.

Supported provider CLIs

Wardian adapts real local provider CLIs into a shared operational view while preserving the provider-specific behavior that matters.

Codex

Real-workspace execution with per-agent habitat state.

Claude Code

Session-aware runtime with permission hooks.

Pi

Real target workspace with explicit project session resume.

Gemini CLI unmaintained

Real-workspace runtime with skill discovery support.

Antigravity

Native AGENTS.md discovery and transcript-based turn state.

OpenCode

Native AGENTS.md discovery with Wardian-scoped config.

Questions

Does my code leave my machine?

Not because of Wardian. It has no server and collects nothing. Your provider CLI connects to whichever model you have configured, on your own account, and Wardian never proxies that request. Point the CLI at a local model and nothing leaves the machine at all.

How is this different from running the CLIs in my own terminals?

Nothing stops you, and Wardian runs the same CLIs in the same workspaces. What it adds is what separate terminals cannot hold between them: agents that can address each other by name, a durable record of what was asked and answered, memory that keeps the evidence it came from, and workflows that branch, loop and wait without you driving each step.

Do I need a Wardian account or a second subscription?

Neither. Wardian runs the CLIs already installed on your machine, signed in to your own accounts, and never holds a key or bills for a request. You pay your model providers exactly what you would without it.

What happens to a running agent if I close its tab?

It keeps running. Closing a tab is a view change, not a lifecycle change; stopping an agent takes an explicit Pause, Restart Session or Delete, so nothing dies because a window closed.

How finished is it?

Early, and under active development. CLIs, on-disk formats and layouts can still change between releases, so pin a version if you depend on it and please file an issue when something breaks.

Downloads

Pre-built binaries are published on GitHub Releases. Choose the asset that matches your operating system and CPU.

Windows Standard x64 installer. Download macOS Apple Silicon For M-series Macs. Download macOS Intel For older Intel Macs. Download Linux Debian/Ubuntu Installable x64 Debian package. Download Linux AppImage Portable x64 Linux app. Download

macOS builds are Developer ID signed and notarized by Apple. Open the DMG, drag Wardian to Applications, then launch it from there rather than from the mounted image. On Windows, SmartScreen can show a reputation warning for a brand-new release.