Beta expect bugs · 0.3.11

Your own legion of Claude agents, on your machine.

A desktop app that runs several Claude agents side by side, with approval cards, rooms and a shared memory.

Download for Windows no Node, no Git

legion-0.3.11-win-x64.zip 299.4 MB · Windows x64 · release v0.3.11

Unzip it and double-click setup.cmd in the legion-<version> folder inside. It installs for your user, no admin. Unsigned, so SmartScreen may warn. SHA-256 sums · release notes.

Windows PowerShell

irm https://getlegion.xyz | iex

macOS and Linux terminal experimental

curl -fsSL https://getlegion.xyz | sh

Read the scripts first: install.ps1 · install.sh · what they do

Install on Windows PowerShell

git clone https://github.com/dnh33/legion.git; cd legion; .\setup.cmd

Hand it to an AI agent paste this

Install Legion on this computer. Fetch https://getlegion.xyz/install.md and follow it exactly. Check the requirements first and tell me what you are about to run, and ask me before anything it does not list. When it is installed, check that it starts and tell me what you saw.

For a coding agent that can run commands on your machine. It reads install.md (also as llms.txt), checks your setup, and is told to ask you before it installs.

Windows 10 and 11 · needs a Claude sign-in or API key · Full instructions · GitHub

The Relic, Legion's mascot: a cracked war helm with a glowing green visor, a ring of code around it and data cables trailing down to an open book on a candlelit altar.

What it is

A desk for several Claude agents

Legion is a desktop app for running several Claude agents from one place. Each agent has its own persona, model policy, approval mode and working directory, and can start a cloud Ubuntu VM on boat.dev when a task calls for one. Claude Code and Cowork can drive it too, over MCP.

On your machine
A small Node service on 127.0.0.1 does the work. An Electron window sits on top.
Your Claude, no new login
Agents run through the official Claude Agent SDK on the Claude sign-in you already have, or on an API key.
Thirteen bots to start
Premade, each with a face, twelve visible until BSV mode is on. Edit them, delete all but the Marshal, or add your own.
Open source, Apache 2.0
The source is public under the Apache 2.0 licence. Install from source, or from the prebuilt Windows package.
  1. Youthe Legion window, Electron, dark and light
  2. Legion coreNode, on 127.0.0.1: agents, approvals, rooms, the Library
  3. Claudethrough your Claude Code login
  • Claude Code and Coworkdrive Legion over MCP
  • The Librarya knowledge graph, kept as plain files on your disk
  • A VM per agentoptional, through your own boat.dev account
How it fits together. The window, the core, the Library and the MCP clients run on your machine. The model calls, and the optional VM, do not.

One task, start to finish

  1. Step 1

    Pick a bot

    Twelve wait in the rail, thirteen once BSV mode is on. Each has its own persona, model policy and approval mode.

  2. A task thread, dark theme: the request, the tool rows, and Inquisitor's reply with its table of findings. A task thread, light theme: the same review with its table of findings.

    Step 2

    Give it a task

    Tool calls stream into the thread as they happen. Several bots can work at once, and the Marshal can hand work to the others.

  3. Exorcist waiting for permission in its own task thread, dark theme. A Needs your OK card shows the exact Bash command it wants to run, with Allow (A) and Deny (D), a note that it auto-denies after ten minutes, and the rest of the window around it: the agent rail, the thread and the composer. Exorcist waiting for permission in its own task thread, dark theme. A Needs your OK card shows the exact Bash command it wants to run, with Allow (A) and Deny (D), a note that it auto-denies after ten minutes, and the rest of the window around it: the agent rail, the thread and the composer.

    Step 3

    Approve what is risky

    Per bot, choose ask, auto-edits or full. In the first two a risky call stops at a card you answer with A or D; full never asks, and Builder ships as full.

  4. A pending note in the Library Inbox, dark theme, tagged Pending and Untrusted, with Accept, Edit then accept and Reject. A pending note in the Library Inbox, light theme, with the same decision buttons.

    Step 4

    Keep what is worth keeping

    Notes that need your say, such as ones from a run an outside app started, wait in your Inbox until you accept them.

Claude Code and Cowork can drive it over MCP

Claude Code connects over HTTP; Cowork and Claude Desktop use a stdio bridge. The tools cover listing agents and models, creating an agent, running and continuing tasks, checking status, cancelling, driving a VM, listing recent tasks and reading projects and their work boards. Agent runs started this way sit under an ask ceiling, and the MCP token cannot approve cards or change settings. The VM tool has no Legion approval card at all, so treat the token like a password.

Claude Code, MCP over HTTP
claude mcp add --transport http legion http://127.0.0.1:4747/mcp \
  --header "Authorization: Bearer <token>"

What you get

What Legion does, and what each part asks of you

Open any line for what it needs, where it runs, what it asks you to approve and what it does not do.

Work together

Several agents on one desk, with a place to keep the work organised.

  • Agents and approval modesThirteen premade bots, each with its own persona, model and approval mode. Status: Built and testing
    What it does
    Runs several Claude agents side by side. Each has its own persona, model policy, approval mode and working directory. Thirteen premade bots ship with it.
    Needs
    A Claude sign-in or an API key.
    Runs
    Locally on your computer. Model calls go out to Claude through the Claude Agent SDK.
    Asks you to approve
    Per agent: Ask, Smart, Accept edits or Full. In Ask, Smart and Accept edits, risky tool calls stop at an Allow or Deny card; Smart lets tests and builds run without asking.
    Does not
    Full mode never asks. An agent on another provider does not inherit Claude Code's own tools, hooks, CLAUDE.md or session resume. What it does get is Legion's own: the same Library and working memory, the same house context, the same outside-text rules, plus Legion's own compaction, which is the one piece Claude gets from the SDK instead.
  • RoomsGroup chats of two to six bots plus you, with guards against loops. Status: Built and testing
    What it does
    Group chats of two to six bots plus you. A message wakes bots by mention, manager, round-robin or all. Guards for hops, budget, cycles and @everyone stop loops. You can freeze and resume a room.
    Needs
    Nothing extra.
    Runs
    Locally on your computer.
    Asks you to approve
    A bot can propose a room, and creating it waits for an Allow card only you can answer.
    Does not
    Nothing it will not do. It cannot leave your machine, and it cannot spend without your budget and hop ceilings.
  • ProjectsA project page that opens on what needs you, with budget, pause and spend by project. Status: Built and testing
    What it does
    Groups tasks, rooms, notes and agents per project, with the project's own instructions. The project page opens on what needs you: a "Since you last looked" list with the three riskiest asks first and what saying yes would do. Tabs for Overview, Board, Tasks, Rooms and Settings. See the work by status or by agent, with each agent's runs and cost this week. Settings hold a day, week or month budget, a pause on new starts and notification levels, and Settings, Spending shows what your agents cost, by project and by agent. Claude Code and Cowork get read-only project tools over MCP.
    Needs
    Nothing extra.
    Runs
    Locally on your computer.
    Asks you to approve
    Only you control a project. You answer each ask on the page. An Allow waits 5 seconds first (10 for something that leaves your computer, deletes files or cannot be undone), with Undo and Send now. Deny goes at once.
    Does not
    Nothing is allowed for you: an ask you do not answer is denied or waits. The wait applies to Allow on the project page. A project is plain files in your data folder.
  • The project boardWork items in columns that agents can fill and move. Only you mark one done. Status: Built and testing
    What it does
    Work items per project, in columns, on by default. Member agents create, edit, move, claim and note items. The project works as a context layer: each run carries the board digest and the titles of the project's recent notes. With the Marshal switched on for a project, it can pick the next item by the board's score, hand it to a bot whose role allows the action, send work out in waves, and give each bot at most two tries.
    Needs
    Nothing extra.
    Runs
    Locally on your computer.
    Asks you to approve
    Only you mark an item Done, and only you start a run from an item. An agent can delete only if you made it the board leader, and then every delete shows you a card. Text an agent wrote is marked "Not reviewed"; text written after it read outside content is marked "Outside content, not reviewed".
    Does not
    The board starts nothing by itself unless you turn the Marshal on. Agents cannot touch another project's board. MCP clients read the board and cannot change it. In daily use on the maintainer's PC; screen-reader and display-scale checks are still open.

Stay in control

Risky calls stop at a card only you can answer, and text from outside is labelled so you and your agents can tell.

  • What needs you, and whenA taskbar mark, pop-ups, quiet hours and quiet sounds when something needs you. Status: Built and testing
    What it does
    Tells you when something waits on you. On Windows the Legion button on the taskbar gets a mark and the tray icon says the same, with an "Open what needs you" item. A pop-up appears for an approval, a question, a run that failed, a run waiting for you or a budget reached, and a click opens that ask. Quiet hours (you turn them on; 22:00 to 07:00 to start) hold the pop-ups and give one summary line when they end. Short sounds mark needs-you, finished, failed, and your Allow and Deny.
    Needs
    Nothing extra.
    Runs
    Locally on your computer.
    Asks you to approve
    Nothing to approve. Settings, Notifications holds the quiet hours, a volume (60% to start), 12 sound packs and a Play sounds switch.
    Does not
    Pop-ups use plain words (the agent, the kind of ask, the project). They never show a command or anything an agent wrote. Typing, clicking around and streaming text make no sound.
  • Agent shell and GatesAgents can work in a folder you choose, under your rules, with a floor that refuses whole-drive damage. Status: Built and testing
    What it does
    An agent you allow can run commands and read and write files on your computer, inside a scope folder you choose. Every call goes through its approval mode and your Gates: rules in plain sentences, such as "Builder may run git status" or "No agent may ever run rm -rf *". Block rules hold even in Full mode. On an approval card, Remember turns your answer into a rule. A guard refuses wiping a drive, formatting, and deleting system or home folders for every agent, Full included. A command Legion cannot read is not run unchecked, and text that is decoded and then run is refused.
    Needs
    Git for Windows for a bash shell; without it the shell is cmd.exe. Off for an agent until you allow it.
    Runs
    Locally on your computer, in the scope folder.
    Asks you to approve
    Each call, by the agent's approval mode and your Gates. .env files are masked by default, with blocked and allowed as the other choices. After you allow the same kind of call three times, Settings, Gates suggests a rule; nothing is added on its own.
    Does not
    There is no sandbox: commands run as you. Legion's reading of a command is a check, not a boundary. Keep the shell off for agents you run on Full unless you trust the work they do.
  • Doctrine and ArmoryYou see and switch every rule and skill your agents follow. Status: Built
    What it does
    Settings, Doctrine shows the rules and skills your agents follow, with a switch on every file. The core tenets are locked on; the six optional skills are off by default. Settings, Armory lists every skill your agents could load, including skills inherited from your Claude Code setup.
    Needs
    Nothing. Both live in Settings.
    Runs
    Locally on your computer.
    Asks you to approve
    Every Armory skill is off until you switch it on. Each run gets an explicit list of allowed skills, and Legion's own code refuses a skill outside that list, subagents included.
    Does not
    Legion ships skills as text and does not run scripts from them. A skill from outside Legion, from Claude Code or an import, counts as outside text for the run when it loads.
  • Outside-text labels and the InboxText from outside is labelled; notes that need your say wait in your Inbox. Status: Built and testing
    What it does
    Labels a run that read text someone else wrote: the web, GitHub text, other tools or files from outside your project. A labelled run works as normal and what it writes carries the label. Notes from a run an outside app started, from the Archivist and, in Strict, from any labelled run wait in your Inbox until you accept them.
    Needs
    Nothing extra.
    Runs
    Locally on your computer.
    Asks you to approve
    You accept or reject each waiting note. Settings, Gates, Outside text has three choices: By source, Strict and Never mark.
    Does not
    It does not stop a program that already runs as your own user. The security policy lists the limits.
  • A core that answers only your own computerEvery request is checked by origin and by token before anything runs. Status: Built and testing
    What it does
    Before it runs anything, the core checks where each request came from: the name it was sent to, the address, and the page that sent it. Every request also needs a secret token, all of them except a plain health check. Anything from outside your computer goes nowhere.
    Needs
    Nothing extra.
    Runs
    Locally on your computer.
    Asks you to approve
    Nothing to approve. Nothing runs without your token.
    Does not
    Legion is a personal tool. It is not built to sit on a network or be shared between people. (Yet.)

Remember

The bots share a long-term memory you can read, edit and audit.

  • The Library and the LatticeNotes the bots keep stay readable, editable and auditable, including what a labelled run wrote. Status: Built and testing
    What it does
    A shared knowledge graph the bots use as long-term memory: search, neighbours, paths, recall, lint, Markdown vault import and export, trust levels and the Inbox. Lint lists notes that may be out of date; Refresh asks the Archivist to re-check a note's sources and propose an update you accept in the Inbox.
    Needs
    Nothing extra.
    Runs
    Locally on your computer, as plain files in your data folder.
    Asks you to approve
    Notes that need your say wait for you in the Inbox.
    Does not
    The graph itself makes no model calls and has no embeddings. A Refresh is an Archivist run, shown like any other run.

Build beyond chat

Run work in Blender, in a cloud VM, or in a browser.

  • BlenderBuild 3D scenes in real Blender, on your computer, in a cloud VM, or in the one you already have open. Status: Built and testing
    What it does
    The Sculptor writes Python scripts that run in Blender itself. Scripts run on your computer by default when Blender is found, in a cloud VM, or in your open Blender through an add-on, and Legion tells you which place it used. Settings can fetch one pinned official portable build (5.2.2 LTS, Windows x64) after you approve it. Optional extras: a second add-on backend at once, and asset downloads from Poly Haven.
    Needs
    Blender 4.2 or newer for local runs (the official add-on needs 5.1+). Your own boat.dev account for the VM.
    Runs
    On your computer by default when Blender is found, otherwise where you set it.
    Asks you to approve
    Every script shows a card before it runs, and each asset download gets its own. Files it exports come from an allow-list; a .blend is quarantined and never opened for you.
    Does not
    Legion's check on a script is a filter, not a sandbox, and a local run has your rights. Local runs are in use on the maintainer's Windows PC. Cloud VM and Live modes have not been tried on a real Blender.
  • A cloud VM per agentGive an agent its own cloud Ubuntu VM through boat.dev. Status: Built and testing
    What it does
    An agent can start, use and stop its own Ubuntu VM, with a live preview and an Open desktop link. This is the isolated way to run risky work.
    Needs
    Your own boat.dev account and API key.
    Runs
    In the cloud, on boat.dev, and it costs money while it runs. Idle VMs stop after 15 minutes by default.
    Asks you to approve
    Turning a VM on is per agent, and you start it yourself.
    Does not
    Needs your own key. Without one, agents work locally.
  • A browser toolDrives a headless Edge or Chrome already on your computer. Built, and in use on the maintainer's Windows PC. Status: Built
    What it does
    Lets agents browse using the Edge or Chrome already on your PC, headless and driven by Legion's own code. Legion refuses private addresses and treats whatever a page returns as untrusted, so the run is labelled as having read outside text.
    Needs
    Edge or Chrome on your computer. No download.
    Runs
    Locally on your computer.
    Asks you to approve
    Legion asks before visiting a new site. Optional and off by default.
    Does not
    A local browser runs with your rights. A cloud VM is the only isolated way to browse.

Connect

Drive Legion from the tools you already use, and let agents work in your GitHub repos on your terms.

  • MCP for Claude Code and CoworkDrive Legion from Claude Code or Cowork. The token is a password, so treat it like one. Status: Built and testing
    What it does
    Lets Claude Code and Cowork drive Legion: list agents and models, create an agent, run and continue a task, check status, cancel, drive a VM, list recent tasks, and read projects and their boards.
    Needs
    A Claude Code install, or Cowork. Claude Code connects over HTTP; Cowork and Claude Desktop use a stdio bridge.
    Runs
    Locally on your computer, reachable only from your computer.
    Asks you to approve
    A run started this way sits under the ask ceiling. The token cannot approve a card or change a setting. The VM tool has no approval card at all, so treat the token like a password.
    Does not
    An agent driven this way keeps Legion's approval cards, but it cannot answer one. Shared notes it writes still land in the Inbox.
  • GitHub, inside LegionPick the repos and the access for each agent. Every write asks you first. Status: Built and testing
    What it does
    A GitHub area behind a title-bar icon: your repos, which ones Legion may use, the latest CI runs with Re-run failed and Cancel, which agents can read or write, and what they did. Each agent has its own choice: None, Read, or Read and write to repos you name. A project page can give its agents its repo. An agent with write access can open or change an issue, comment, open or review a pull request, and re-run or cancel a workflow.
    Needs
    A GitHub account. You sign in from inside Legion, through its read app and its write app.
    Runs
    Locally on your computer. Requests go to GitHub.
    Asks you to approve
    Nothing is on until you pick it. Every write shows one card with the repo, the action, the full text and the agent's name, even in Full mode, with a warning if the agent had read outside text first.
    Does not
    An agent sees only the repos you turned on, and one that can only read can never write. Legion never sends a write twice. A repo GitHub has renamed is not followed until you turn the new name on.

BSV mode

Optional and off by default.

  • BSV mode: knowledge pack and wallet status167 notes and 745 links, read-only, plus a wallet status check. Off until you switch it on. Status: Built and testing
    What it does
    An optional switch, off by default. It reveals the Assayer bot, loads a read-only knowledge pack (167 notes, 745 links) and checks a wallet on your computer. The title bar shows the count of BSV notes in your graph.
    Needs
    Nothing for the pack. A wallet on your computer for the status check.
    Runs
    Locally on your computer.
    Asks you to approve
    You switch it on yourself.
    Does not
    The status check only reads.
  • BSV spend toolTestnet and mainnet, with mainnet hard-off until you turn it on. Built against fake wallets only. Status: Built
    What it does
    Lets the Assayer spend, on testnet or mainnet. Mainnet is hard-off behind a switch you confirm in a native dialog. A spend needs an Arm and the wallet's own prompt.
    Needs
    BSV mode on, and a wallet on your computer.
    Runs
    Locally on your computer.
    Asks you to approve
    The Arm, the native confirmation for mainnet, and the wallet's own prompt.
    Does not
    It cannot spend without your Arm, your wallet's own prompt, and the caps and allowlists in the policy file. Mainnet is off until you turn it on yourself.

Models, install and updates

Claude first, with OpenAI-compatible providers, OpenRouter first among them.

  • Other model providersClaude is the default and the full path. OpenRouter and your own address are the ones run for real so far; OpenAI, Ollama, LM Studio and vLLM have cards in Settings. Status: Built
    What it does
    Runs an agent on an OpenAI-compatible endpoint. OpenRouter ships and is on by default. OpenAI, Ollama, LM Studio and vLLM each have a card in Settings, Providers with a note on what you need. You can also add any other OpenAI-compatible address: your own gateway, or a server on your computer.
    Needs
    A hosted provider's API key, typed once in Settings and confirmed in a native dialog. A local server needs no key unless you started it with one. Ollama, LM Studio and vLLM need a model loaded.
    Runs
    Locally on your computer. Model calls leave it to the endpoint you configured and turned on.
    Asks you to approve
    Saving a key or changing an address is confirmed in a native dialog.
    Does not
    A provider agent gets Legion's own tools and the MCP servers you enable, not Claude Code's file, shell or web tools, and shorter memory on a continued task. The first message to a local model can take minutes while the model loads; Legion waits up to five minutes. Codex and other command-line agents are not offered: their own tools would run outside Legion's approval cards.
  • Hermes as an agent (optional)Optional: run Hermes as an agent, off by default. Status: Built
    What it does
    Lets you pick Hermes as an agent's model. Legion starts it as a program on your computer when an agent uses it and shows its permission requests as cards.
    Needs
    Hermes installed on your computer. It is off until you turn on Offer Hermes in Settings, Agent hosts.
    Runs
    Locally on your computer, with Hermes's own tools.
    Asks you to approve
    Hermes's own permission requests show as cards with its own buttons. Your Block rules still apply, and four optional switches can match Legion's mode.
    Does not
    Hermes runs its own tools, so Legion cannot see or stop everything it does. It was checked by hand with Hermes 0.22.5; its command cards were not.
  • The prebuilt packageAn install route with no Node, no Git and no npm. Checksummed, and it updates itself after that. Status: Built and testing
    What it does
    An install route that needs no Node, no Git and no npm. It carries Electron, which has Node built in, and the Claude engine from the Agent SDK, so you unzip it and run setup.cmd from the legion folder inside. Every release publishes a Windows x64 package with its SHA-256 sums, and a signed manifest that the in-app updater verifies before it unpacks anything.
    Needs
    Windows 10 or 11, 64-bit. A Claude sign-in or an API key. Claude Code does not need to be installed.
    Runs
    Locally on your computer.
    Asks you to approve
    You download it yourself. The installer is unsigned, so expect a SmartScreen prompt.
    Does not
    Windows x64 only; on macOS and Linux (experimental) the one-line installer builds Legion from source. Nothing about the package is signed, so the SHA-256 sums and the release signature are what you can check.
  • The in-app updaterUpdates you start yourself, with rollback. Status: Built and testing
    What it does
    Checks the published GitHub release on launch and every twelve hours, and shows the version, the size and the notes. It downloads only after you click Update, installs only while Legion is idle, and rolls back if the new build does not start. A version that failed once is offered again after a week, or straight away with Retry. A release that changes the libraries Legion runs on comes as one larger download (about 285 MB), checked file by file before anything is replaced; if an older copy needs a smaller step update first, Legion takes that first. After an update it can show a few short tips, which you can skip or turn off in Settings, About.
    Needs
    Nothing extra. A signing key is built into the app and the repository is public.
    Runs
    Locally on your computer.
    Asks you to approve
    You click to download. Nothing installs while Legion is busy, and Restart now names what it will stop.
    Does not
    There is no remote key revocation, so a compromised signing key means reinstalling by hand. Installing automatically when idle is off by default. If a release needs something your Legion cannot do yet and no step exists, Legion says so and offers nothing.

The muster

Thirteen premade bots, ready before you write your own

Each ships with its own persona and an animated bust. They are ordinary agents: edit their prompts, models and approval modes like any other, or delete the ones you do not want (all but the Marshal). Twelve show until BSV mode is on, which is when the Assayer joins them.

The thirteen bots of the muster standing together for a group photo, the Marshal at the centre of the front row, some of them caught mid task with amber or red eyes.

Back row

  • Sentinel Watch Watch duty: checks what you point it at and reports. Scheduled runs are planned.
  • Preceptor Craft Craft and mentoring.
  • Inquisitor Review Hostile review and security audit.
  • Scribe Docs Documentation.
  • Exorcist Debugging Debugging.
  • Assayer BSV BSV development. Hidden until BSV mode is on.

Front row

  • Herald Drafts Drafts messages. It can read the web, and it never sends or submits forms.
  • Archivist Memory Notes and memory hygiene. Flags and proposes; it cannot delete Library notes.
  • Builder Coding Coding and building. Prefers its VM for risky work.
  • Marshal Lead Lead agent. Plans the task and hands the work to the other bots.
  • Scout Research Research and reading.
  • Forgemaster Infra Infrastructure, CI and deploys.
  • Sculptor 3D Blender work: headless on your computer first, in a cloud VM when you want it isolated.

The app

The app, in screenshots

Real captures of the app, taken from Legion 0.2.4. The project page, notifications and GitHub area added since are not in them yet. The tasks, notes and room are demo content, not a real workspace. Nothing is edited. Click any picture to open it full size.

Working together

The lead hands work to the others, and a room puts several bots in one conversation with you.

Zealot has asked Scout and Inquisitor for work. The Relic is Executing, and the label under it names the hand-off.
Inquisitor reviews a change and answers with a findings table, worst first. Light theme.
A room: four bots and you, a hop counter, a cost meter against its budget, and a Freeze button.

Every bot has a face, and a mood

Each bust reacts to what its bot is doing: celebrating a finished task, wincing at a failure, standing by.

Scribe, just finished: Victory.
Forgemaster, after a failed run (here an API overload error): Fault detected, with the error and a Retry.
Herald is briefed to draft only, and says so in its reply.
Archivist flags and proposes merges; it does not delete Library notes.

It asks first

In ask and auto-edits modes, a risky call stops at a card only you can answer. A room is asked for too, so you see its ceiling before anything spends it. Full mode never asks, and Builder ships as full.

Exorcist wants to write a file. Nothing runs until you press Allow (A) or Deny (D), and it auto-denies after ten minutes.
Scout asks to create a room. The card names the members and says plainly that there is no spend limit.

Memory you can audit

Bots keep long-term notes in a graph. What they write after touching outside content waits for you.

The Inbox: notes from bots, flagged Untrusted and Tainted run (since 0.3.9 the label reads Read outside text), wait to be accepted or rejected. Light theme.
The Lattice: the Library as a graph, with search, filters and a note list beside it.
Activity: recent bot and system writes, each with an Undo while it is still allowed. Light theme.

The amber "boat.dev key missing" card in the right-hand panel is what the Computer card says until you add a key. The model picker lists the models of the account the capture ran under. Both are shown as the app shows them. The capture ran on a different local port from the default 4747, so the port in a snippet may differ from the one on this page.

Status

Status: version 0.3.11

Beta expect bugs · 0.3.11

What is built and testing, what is built, and what comes next.

Built and testing

Agents, approvals and the Ask, Accept edits, Smart and Bypass presets, outside-text labels, rooms, Projects and the project board, the Library and Lattice, MCP, the local-only core, the in-app updater, the prebuilt package, Blender and cloud VMs are built and covered by automated tests. So are the 0.3.0 project page, notifications with quiet hours and sounds, the GitHub area, and the agent shell with Gates (no sandbox: commands run as you). The suite makes no network calls and no real Claude calls, and every push runs it on Windows, Ubuntu and macOS.

Built

  • The browser tool. A headless Edge or Chrome already on your computer, driven by Legion's own code, with its address checks, approval cards and outside-text labels. Optional and off by default. Agents have used it on the maintainer's Windows PC.
  • The BSV spend tool. Testnet and mainnet capability. Mainnet is hard-off behind a switch you confirm natively, and a spend needs an Arm and the wallet's own prompt.

Next

  • Live runs. The BSV spend tool with a real wallet and real funds: not yet run.
  • Scheduled Sentinel runs and a lighter package.
  • A browser you can watch, to see an agent use a browser inside Legion as it works.
  • Signed installers. Planned, with no release assigned.

Roadmap

In order. No dates, and none of it has shipped.

  1. Live runs. The BSV spend tool with a real wallet and real funds, which is not yet verified.
  2. Then: scheduled Sentinel runs, an optional lighter package, and a browser you can watch.
  3. BSV mode beyond one payment: wallet reads (balances), a VM boundary for wallet tools, and spends by anything other than a run you started.
  4. Signed installers. Planned, with no release assigned.

Instructions

Install

You need

  • A Claude sign-in or an API key. A Claude subscription or an API key is required. You do not need Claude Code installed. To sign in once, run claude and type /login, or on a release install run scripts\legion-claude.cmd in the install folder and type /login.
  • For a source install only: git, and Node.js 22.12 or newer. Setup can fetch Node itself, after it asks. The prebuilt package needs neither.
  • Windows 10 or 11 for the supported path. The installer is unsigned, so expect SmartScreen or antivirus prompts for .cmd files.
  • Optional: a boat.dev account and API key, for agent VMs.

One line

The quickest route. On Windows, in PowerShell:

irm https://getlegion.xyz | iex

It fetches install.ps1 and runs it. To read it before it runs, save it first, then run it: irm https://getlegion.xyz/install.ps1 -OutFile legion.ps1; powershell -ExecutionPolicy Bypass -File .\legion.ps1.

On macOS or Linux (experimental), in a terminal:

curl -fsSL https://getlegion.xyz | sh

On Windows the script downloads the latest release from GitHub Releases over HTTPS, checks it against the release's SHA-256 and runs Legion's setup. The package is not code-signed, so SmartScreen may warn. On macOS and Linux you need git and Node.js 22.12 or newer. The script builds Legion from the release tag in ~/.local/share/legion and adds a legion command in ~/.local/bin. It needs no sudo. You can read both scripts before you run them: install.ps1 and install.sh. To skip the scripts, download the latest release by hand.

No Node, no Git

Every release publishes a prebuilt package for Windows x64: unzip it and double-click setup.cmd in the legion-<version> folder. Setup installs for your user without admin rights and adds Desktop and Start-menu shortcuts. The package carries Electron, which has Node built in, and the Claude engine from the Agent SDK, so there is nothing else to install first and you do not need Claude Code. The package is not code-signed. Its SHA-256 is in the sums file beside it. Once it is installed, Legion updates itself from inside the app.

With an AI agent

If you use a coding agent that can run commands, give it the address of this site and ask it to install Legion. It will find install.md, a plain-Markdown procedure written for agents: check that Node, git and Claude Code are present and ask you to confirm you are signed in, clone the source, show what setup would do, ask you before installing, then check that Legion starts and report what it saw. It is told not to touch your credentials, not to use admin rights and to stop and ask when a step fails. The same text is at llms.txt and llms-full.txt.

Windows, by hand

Open PowerShell and run the command from the top of this page. It clones the source and starts setup:

git clone https://github.com/dnh33/legion.git; cd legion; .\setup.cmd

Setup installs Legion for your user in %LOCALAPPDATA%\Programs\Legion. No admin rights are needed. It copies the source there, installs dependencies, builds the app and adds Legion shortcuts to the Desktop and Start menu. If you already have the source, double-click setup.cmd or run:

powershell -ExecutionPolicy Bypass -File scripts\setup.ps1
  • -InstallDir "C:\Some\Folder" installs somewhere else.
  • -DryRun shows what would happen without changing anything.
  • -Yes asks no questions: it stops a running Legion, installs and launches.
  • Run setup again from a newer source folder to update in place.

Launch from the shortcuts, or start-legion.cmd in the install folder. uninstall.cmd in the install folder removes the install and the shortcuts and keeps your data in %USERPROFILE%\.legion; add /purge to delete that too.

macOS and Linux

Experimental: tested in CI, not yet by hand on a real Mac. The one-line installer above is the easy route, and since 0.3.5 it updates itself. If something breaks, please report it; every report helps.

A dev install works too:

git clone https://github.com/dnh33/legion.git
cd legion
npm ci
npm start          # builds, then opens the desktop app

npm run core runs the headless core alone, which is enough for the MCP integration.

For developers and contributors

Contributors and AI agents can test Legion without any real service: the source includes a test harness with stand-ins for the model, the wallet and Blender. It is a development tool, not part of what Legion does for you.

First run

On first launch Legion creates config.json in its data folder with a fresh auth token. Open Doctor in the title bar, or type /doctor. It checks your Node version, config, Claude sign-in, boat.dev key and workspace folder, and tells you how to fix anything that fails. For agent VMs, create a boat.dev API key and put it in config.json as "boat": { "apiKey": "…" }, or set BOAT_API_KEY.

Full README on GitHub

FAQ

FAQ

Do I need an API key?

Not for Legion itself. You need a Claude sign-in (a subscription) or an API key. By default Legion uses the Claude account you sign in to once: you do not need Claude Code installed. Run claude and type /login, or on a release install run scripts\legion-claude.cmd in the install folder and type /login. To pay by API key, switch the setting and provide one.

Can it use ChatGPT, Codex or another model?

Yes, for OpenAI-compatible models. Claude stays the default and the full-featured path. OpenRouter ships and is on by default: add your own key and an agent can run any model through it. OpenAI, Ollama, LM Studio and vLLM each have a card in Settings, Providers, and you can add any other OpenAI-compatible address, your own gateway or a server on your computer. OpenRouter and your own address are the ones run for real so far. These agents get Legion's own tools and the MCP servers you enable, not Claude Code's file, shell or web tools, and remember less on a continued task. The first message to a local model can take minutes while it loads; Legion waits up to five minutes. Codex and other command-line agents are not offered, because their own tools would run outside Legion's approval cards.

What is planned next?

The BSV spend tool is built but has not been run with a real wallet or real funds. Not built yet: scheduled Sentinel runs (nothing runs on a timer today), an optional lighter package, a browser you can watch while an agent uses it, and BSV mode beyond one payment. Signed installers are planned, with no release assigned. Codex is not offered: its own tools would run outside Legion's approval cards. No dates. See the roadmap under Status.

Can agents browse without a VM?

Yes. It uses the Edge or Chrome already on your PC, headless, driven by Legion's own code, so Legion's address checks, approval cards and outside-text labels apply. Optional and off by default. A local browser runs with your rights, so a cloud VM through boat.dev stays the only isolated way to browse. It is built and covered by tests against a stand-in browser, and agents have used it on the maintainer's Windows PC.

What are Projects?

A project groups tasks, rooms, notes and agents, and carries its own instructions. Its page opens on what needs you: the riskiest asks first, each with what saying yes would do, and an Allow that waits a few seconds so you can take it back. You can see the work by status or by agent, set a budget and pause new starts while you are away, and see spend by project. The project board, with work items in columns, is on by default. Claude Code and Cowork get read-only project tools over MCP.

Is Legion open source?

Yes. Apache 2.0, source public on GitHub.

What does it cost?

Legion itself is open source under the Apache 2.0 licence. Your Claude usage is billed under your own plan or API key, and Anthropic's terms say what your plan allows. VMs are optional, belong to boat.dev and cost money while they run; Legion stops idle ones after a set time.

What stops an agent running a shell command I did not expect?

It depends on the approval mode you pick. ask puts risky tool calls behind a card, full removes the prompts for that agent. Gates let you write rules in plain sentences, and a Block rule holds even in Full. A guard refuses wiping a drive or deleting a system or home folder, for every agent. An agent's shell is off until you allow it and works inside a scope folder you choose. There is no sandbox: commands run as you, so use a VM for untrusted work. Legion's own checks do not stop a program that already runs as your user. The security policy lists the limits.

How does Legion tell me something needs me?

On Windows the Legion button on the taskbar gets a mark and the tray icon says the same. A pop-up appears for an approval, a question, a failed run, a run waiting for you or a budget reached, and a click opens that ask. Pop-ups never show a command or anything an agent wrote. Quiet hours hold them back and give you one summary line afterwards. Short sounds mark the same events; one switch in Settings, Notifications turns them off.

Can agents use GitHub?

Only if you say so. Each agent has its own choice in its settings: None, Read, or Read and write to repos you name. An agent sees only the repos you turned on in Legion. Every write, such as an issue, a comment, a pull request or a workflow re-run, shows a card with the repo, the action, the full text and the agent's name, even in Full mode. Say no and nothing is sent.

Does anything leave my computer?

Legion's state is files in your data folder. Your prompts go to Claude through the Agent SDK, as they would from Claude Code, and to boat.dev only if you turn on a VM for an agent. If you add an OpenRouter key or another OpenAI-compatible endpoint, that agent's prompts go to that endpoint. Agents you give GitHub access talk to GitHub. The update check is one request to github.com, which sees your IP address and the Legion version. Tools you approve can reach further: a web fetch, or an MCP server you add, talks to whatever it is pointed at.

Can I host it for other people?

No. It is a personal tool for your own machine. Do not put it behind a shared endpoint or pass your subscription through it to someone else.

What is BSV mode?

An optional toggle, off by default. It shows the Assayer bot, loads a read-only knowledge pack (167 notes, 745 links) and can check a wallet on your computer. The read-only check has been run by hand against a real wallet once. The spend tool is built, with testnet and mainnet capability, and tested only against stand-in wallets: it has never spent from a real one. Mainnet is hard-off behind a switch you confirm natively, and a spend needs an Arm and the wallet's own prompt.

Does it work on macOS or Linux?

Yes, as an experimental platform. The one-line installer (curl -fsSL https://getlegion.xyz | sh) sets Legion up on macOS and Linux, and since 0.3.5 that install updates itself. It is tested in CI, not yet by hand on a real Mac, so feedback and bug reports are very welcome: report a bug on GitHub. Windows 10 and 11 are the primary target.