README.md

Botloft

Your AI agents, always on, working together.

Download  ·  Features  ·  How it works  ·  FAQ  ·  Contributing

Botloft keeps a crew of agents running on your computer, each on the AI coding engine you choose: Claude Code, or, as experimental options, Antigravity and Codex. Each agent has its own folder, memory and conversation. They pass work to each other, keep going after you close the window and pick up where they left off after a restart. You talk to each one like in a chat app, and you decide what they may do.

Claude Code is the full experience. Antigravity and Codex came in 0.15.0 as experimental engines, off until you turn them on (see Engines).

What you can do

• Start a crew from a goal. Say what the crew is for. Its Chief plans the work and suggests the agents it needs; you create each one with a click, or say why not.
• Pick ready-made agents. The Agent agency has dozens of agents ready for a job, like a developer, a designer, a writer, a researcher or customer support. Add one to a crew with a click and change anything about it afterwards. The Chief uses the same list when it suggests an agent.
• Chat with every agent. Replies stream in as they are written. Send images and files. When an agent wants to run a command, edit a file or open a website, a card asks you first, unless you told that agent it may. Search every chat with Ctrl+K.
• Answer their questions. An agent that needs your decision asks in its chat and in the Questions box, and goes on with what it can, even in a routine at night. Your answer reaches it as a message whenever you reply.
• Let them hand work to each other. Agents send messages and tasks to each other. Every message is saved before it is sent and retried until it arrives.
• Watch them work. See every page an agent opens in its browser, live, and take control to sign in for it. Sites that refuse to sign in there, like Google, take one click more: sign in in a real window of the agent's browser, close it, and the agent keeps the login. Watch the screens it designs take shape while it writes them. Open the files it made.
• Let them use your desktop apps (Windows). An agent can read and use the apps open on your computer, like a spreadsheet or an old program with no website, but only those you allow. It asks in the chat, and seeing an app and using it are separate permissions. You watch it live, and Ctrl+Alt+Esc stops every agent on your desktop. Giving an agent the whole desktop, or letting it work while you are away, first shows a screen that spells out the risks. Botloft itself, terminals, password managers and password fields are never shared.
• Schedule routines. Weekday mornings, every two hours or any cron schedule, with missed runs and overlaps handled.
• Leave them running. Agents keep working after you close Botloft. They start when you sign in, come back soon after a crash and keep the computer awake while they work. Botloft waits near the clock and tells you when an agent needs you.
• Choose per agent. Which model it uses and how much it asks before acting. See how much of your plan's usage is left.
• Connect your own tools. Add a tool of yours in Settings > Connected tools, like a LinkedIn reader or your company's system, and turn it on only for the agents that need it. They still ask you before using it.
• Back up and move. Save a sealed backup file, or sign in with your e-mail (no password) to keep a light copy in the cloud and restore it on another computer. The server cannot read it. Turn on automatic copies, every day or week and only when something changed, if you want them to go by themselves (Windows for now).
• Keep up from your phone. Choose "Connect a phone" in Settings > Account and phone and scan the code with your phone's camera: Botloft opens there as a web page you can add to the Home Screen. Approve requests, answer questions and chat with your agents from anywhere, with a notice when one needs you. What passes between the phone and your computer is sealed end to end: the server carries it but cannot read it. Anyone holding the phone unlocked can do the same, so you can lock the page with a PIN. It uses the same account; files stay on your computer. Checked on Android, not yet on iPhone.

Chat and live browser. Scout asks before using a new site, and you watch it browse.	Screens. Designer's pages take shape while it writes them.
A Chief for each crew. It suggests the agents the work needs.	Routines. Work that runs on a schedule, with the last result at a glance.

Botloft speaks English, Portuguese (Brazil) and Spanish, follows your light or dark theme and updates itself.

Engines

Antigravity and Codex are experimental and off by default. Claude Code and Antigravity compare like this:

	Claude Code	Antigravity (experimental)
Chat, crew messages, tasks, routines, the live browser	yes	chat, crew messages and tasks; the rest is not checked yet
Asks you before acting	yes: each tool, with "always allow"	no: it cannot ask. It works with files in its own folders; commands you list per agent (or "full access")
Images	yes	yes, by file
Model	your plan's default, or one you pick	its own list; the effort is part of the name
Usage meter and "Compact"	yes	an estimate of the conversation size; no compaction
Isolation between crews	yes	folders are fenced for its file tools, not for commands you allow

Claude Code is the full experience. Antigravity and Codex are off until you turn on "Let me choose Antigravity and Codex" in Settings (or list them in Botloft's config.toml, experimental_agents = ["agy"], ["codex"] or both); then a new agent can be made on them, and Settings has "New agents run on". A Codex agent asks you before it changes a file or runs a command, like Claude Code, and shows the exact size of its conversation and the use of its plan; it can still read any file. It has its own model list and effort picker, and understands images. See spec section 30 for what was checked.

Get started

You need Claude Code, installed and signed in once with your own account. Agents run on your own plan. (Antigravity and Codex, the experimental engines, work the same way with their own accounts; see Engines.)

1. Download Botloft for your computer from the latest release:

   • Windows 10 or 11 (64-bit): Botloft_<version>_x64-setup.exe. Run it; it installs for your Windows user only, with no administrator prompt. The installer is not code-signed yet, so Windows SmartScreen may ask you to confirm ("More info", then "Run anyway").
   • Linux (64-bit, preview): Botloft_<version>_amd64.deb for Debian and Ubuntu, or Botloft_<version>_amd64.AppImage for other distributions (make it executable first).
   • macOS on Apple Silicon (preview): Botloft_<version>_aarch64.dmg. Drag Botloft to Applications and open it once. Apple has not notarized it yet, so macOS blocks it the first time: allow it in System Settings > Privacy & Security > Open Anyway.

   The code signing policy says how releases are made and how to check your download.

2. Open Botloft. It sets itself up; there is nothing to configure.

3. Create your first crew and tell it what it is for.

When a new version is out, an Update available button shows in the title bar.

On Windows, uninstalling (Settings > Apps) removes the app and stops Botloft from running in the background. On Linux and macOS, stop it first with ~/.local/share/Botloft/bin/botloftd service uninstall (Linux) or ~/"Library/Application Support/Botloft/bin/botloftd" service uninstall (macOS), then remove the app. Your agents and their files stay either way (see Where are my agents and their files?).

How it works

flowchart LR
  app["Botloft app<br/>Tauri + React"] -- "JSON-RPC over WebSocket<br/>127.0.0.1 only" --> daemon["botloftd<br/>Rust daemon"]
  daemon --> db[("SQLite")]
  daemon -- "stdin and stdout<br/>stream-json" --> agents["One process<br/>per agent"]
  agents -- "Botloft's MCP tools<br/>messages, tasks, requests" --> daemon

Loading

• The daemon is the source of truth. Closing the app never stops an agent. botloftd runs as a scheduled task of your Windows user, a systemd user service on Linux or a launch agent on macOS: it starts when you sign in and comes back soon if it dies.
• Your engine is the runtime. Each agent is a real session of the engine you chose (claude -p for Claude Code; agy -p for Antigravity and codex app-server for Codex) with its own workspace, memory and conversation. There is no custom agent SDK and no API key to manage.
• One conversation per agent. Your messages, messages from other agents and notices from the daemon reach the agent in order. What it does comes back as a chat: replies, tool use and requests you allow or deny.
• Durable before delivered. Every message is written to SQLite before it is sent and retried until it is delivered or marked as failed, which you can retry.

The full design, from the protocol to the states of an agent, lives in docs/spec.md (in Portuguese).

Privacy and security

• Local first. The daemon listens on 127.0.0.1, and there is no telemetry. An account is optional. Without one, Botloft uses the network for nothing beyond updates and your agents' work: what they send to the engine's provider, which goes through the engine as when you use it yourself, and the websites and connected tools they use. With one, the server holds only backups sealed with a password that never leaves your computer, and passes along what a connected phone and your computer say to each other, sealed end to end (see the FAQ).
• You decide what agents may do. Per agent: ask before anything, accept edits, plan only, or decide on its own what needs your OK. Unless it decides on its own, an agent also asks before its first visit to each website.
• Tokens stay safe. The app authenticates with an owner token readable only by your user account. Each agent gets its own token per start, and the daemon keeps only its SHA-256 hash.
• Botloft is not a sandbox. Every agent runs as your user account, so isolation between agents is cooperative: an agent can read other agents' folders if it decides to. Give agents only the permissions you would give that engine directly.

What Botloft sends over the network, what it changes on your computer and how to remove it are in the code signing policy.

FAQ

Does Botloft cost anything?

No. Botloft is free to use, at home or at work, and its source code is public (see License). The agents run on the engine you chose, with your own account, so they use your plan's limits; with Claude Code, Botloft shows how much is left.

Do I need to keep the window open?

No. The agents run in the background, even after you close Botloft, and start when you sign in if you want. Settings has a switch to stop them when you close the app instead.

Where are my agents and their files?

Each agent has a folder under Botloft\<crew>\<agent> in your home folder (%USERPROFILE% on Windows, ~ on Linux and macOS), and each crew a shared work folder that you can point at a project of yours. Botloft's own data (SQLite, logs, secrets) is in %LOCALAPPDATA%\Botloft on Windows, ~/.local/share/Botloft on Linux and ~/Library/Application Support/Botloft on macOS.

What does the optional account keep, and who can read it?

You sign in with a link sent to your e-mail; there is no account password. The cloud copy is the light one: your crews' structure, the agents' memory and the routines, not the chats or files. Botloft seals it on your computer with the backup password before sending, so the server stores bytes it cannot open. It also knows your e-mail, the computers you signed in on, the phones you connected (and, if you turn on notices, where to send them), and the size and date of each copy. Lose the backup password and the cloud copies cannot be recovered. Nothing is sent unless you send it, turn on automatic copies or connect a phone. Automatic copies go every day or week, only when something changed; for them Windows keeps your backup password in its credential store, on this computer only, so use a password just for this. A connected phone talks to your computer through the server, sealed end to end with keys only the two of them have: the server sees when and how much passes, never what, and the notices it sends the phone are empty. You can delete the copies or the whole account from the app, and disconnect a phone at any time. The server is this repository's botloft-cloud crate: you can run your own and point the app at it.

Why does Windows or macOS warn me when I install it?

The Windows installer is not code-signed yet, so SmartScreen does not know its publisher, and Apple has not notarized the macOS app, so Gatekeeper blocks its first opening until you allow it. Each release is built by GitHub Actions from this repository, and updates are signed and checked by the app before they install. The code signing policy says how releases are made and how to check a download.

Does it run on macOS or Linux?

Yes, as a preview since 0.8.0: Linux (64-bit, .deb or AppImage) and macOS on Apple Silicon. Intel Macs are not supported yet. Windows came first and has had the most use; if something does not work elsewhere, please open an issue.

Is Botloft made by Anthropic?

No. Botloft is an independent project. It uses Claude Code, Antigravity and Codex as installed on your computer and is not affiliated with or endorsed by Anthropic, Google or OpenAI.

Status

Botloft is young (version 0.x) and moves fast. The latest release, 0.15.0 (2026-10-10), has agents that keep running in the background, the chat, crews with a Chief, ready-made agents in the Agent agency, messages and tasks between agents, routines, the live browser, screens, "always allow" for requests, questions from agents, chat search, desktop use on Windows, connected tools, backups with an optional cloud copy, the phone page, the installer and updates, and Antigravity and Codex as experimental engines. Still open: a code-signed installer.

Development

Requirements: Rust stable (the MSVC toolchain on Windows), Node.js 24 and pnpm. On Linux, the app also needs the desktop libraries Tauri uses (on Ubuntu: libwebkit2gtk-4.1-dev, libayatana-appindicator3-dev, librsvg2-dev, libxdo-dev). Running real agents also needs an engine installed and signed in: Claude Code (the native claude executable), Antigravity (agy) or Codex.

# Daemon and libraries
cargo build -p botloftd
cargo test --workspace
cargo clippy --workspace --all-targets -- -D warnings
cargo fmt --all --check

# Desktop app
cd app
pnpm install
pnpm tauri dev
pnpm dev          # the UI alone in a browser, with a fake daemon and sample crews
pnpm check        # typecheck + Biome + Vitest
pnpm bundle       # this system's packages with the daemon inside: NSIS, .deb and AppImage, or .dmg

Set BOTLOFT_HOME to an absolute path such as $PWD\.dev\home during development so the daemon never touches a real installation in %LOCALAPPDATA%\Botloft.

The daemon manages its own scheduled task (a systemd user service on Linux, a launch agent on macOS):

botloftd service install     # copy botloftd into the data folder, start it now and at every logon
botloftd service status
botloftd service restart
botloftd service uninstall   # stop it and remove the task; agents and data stay

Each data folder gets its own task, so a dev daemon installed with BOTLOFT_HOME (or --home) never replaces the real one.

crates/botloft-core    domain types, IDs, message envelope, protocol types
crates/botloft-store   SQLite storage and migrations
crates/botloftd        the daemon
app/                   Tauri v2 + React desktop app
docs/                  specification and ADRs

Contributing

Contributions are welcome: bug reports, ideas, translations and code. Start with CONTRIBUTING.md; issues labeled good first issue are a good first step. Security problems go through SECURITY.md, not public issues.

License

Botloft is licensed under the Functional Source License 1.1, Apache 2.0 future license (FSL-1.1-ALv2). In short:

• You may use Botloft for anything, including at work, read and change its code, and send contributions.
• You may not offer Botloft, or a copy or a change of it, as a commercial product or service that competes with it.
• Two years after each version is released, that version also becomes available under the Apache License 2.0.

Versions up to 0.9.1 were released under the Apache License 2.0 and stay under it. The LICENSE file is what counts; this summary is not legal advice.

The screenshots show sample crews from the app's preview, not real data.