Agentbits is a desktop app where you build your own team of AI agents (a lead plus specialists with different
roles) and hand them work the way you would hand it to people. The lead splits your request into tasks, assigns
them, reviews the results and remembers what was decided. You watch the whole team work live in a pixel-art
office, on a Kanban board and in each agent's terminal, all from one window, with your data kept on your
own computer.
At a glance
| |
|---|
| What is it? | A multi-agent workspace: an organization of AI agents with roles, teams and a lead, working on projects you define. |
| What problem does it solve? | Instead of juggling one chat window per task and copy-pasting between them, you describe the goal once. The team plans it, does it in parallel and reports back, and you can see who is doing what at any moment. |
| How do AI teams work here? | You talk to the project's lead. The lead turns your message into tasks, the autopilot hands them to free team members, each agent works with its own permissions, and the lead reviews the result (approve or send back for revision). |
| How is it different from a single assistant? | Several agents work at the same time, each with its own role, job description, skills and access rights. The work is visible (office, board, terminals), reviewed by a lead, and the project keeps a long-term memory and a decision log. |
| Why use it? | One place to run and watch AI work across projects, with real guardrails: deny-by-default access, approval prompts, an audit log, and secrets that never leave the main process. |
| I'm not a developer. Is it for me? | Yes. Ready-made team templates cover research, marketing, product and e-commerce work, and a free Demo mode lets you walk through the whole workflow as a simulation, without an API key. Real work needs an AI provider (see below); some tasks, such as browsing the web or editing files, only work with certain providers. |
Highlights
- Your own AI organization. Companies, departments and teams nested as deep as you like, with agents that
have a role, a job description, skills, a look and a permission preset. Six ready-made templates (Core Team,
Software Team, Research & Analysis, Product Development, Marketing & Content, E-commerce Company), or start from scratch.
- Projects with a lead. Each project is a folder plus the teams that work on it. You chat with the lead, and it
turns the request into tasks on the board. Only the lead talks to the team: it hands out work, sees what members
finished and leaves them notes. Replies appear as they are written, and unsent messages stay as drafts per chat.
Opening a project never starts work by itself: you press Start work in this project.
- Parallel developers. Duplicate an agent into twins; the lead splits work between them, tasks can wait for each
other, and everyone sees who is working on which files. Optionally each task works in its own git working copy and is
merged when approved, so agents never overwrite each other.
- Race mode (optional). A task also goes to a member on another provider or model; the lead compares both attempts
and keeps the better one.
- Share teams. Export a unit or an agent as a
.agentbits file and import one from a file or a GitHub link, with a
full preview first. Keys, connections, tasks and chat never leave.
- Live pixel-art office. Agents sit at their desks while they work, take coffee breaks, and show approval or
error badges. Click an agent to see what it is doing.
- Kanban board. To do → In progress → In review → Done, with drag and drop, per-task activity, logs and runs.
Moving a running task to Done asks first and never merges anything by itself. "Cancel task" stops its run and turns
down its approvals; nothing picks it up again until you reopen it, and its history stays.
- Agent terminals. A persistent transcript for every agent: what it was asked, which tools it used, and token
counts. You can redirect an agent, queue an instruction or give it a new task from there. Real interactive shells
(PowerShell / bash) are available too.
- Needs you. One list, in the top bar, of everything waiting for you: approvals, unmerged work, reviews the lead
couldn't decide, held tasks with their reason. Cards say why a task doesn't move. A failed task is held at once,
with a plain-language reason, until you retry it. The narrow exception (Settings → General → "Automatic recovery"):
after a passing provider failure (network, timeout, overload), a lead review on an API provider or Demo is tried
again, and a worker run on an API provider or Demo is run again only when its record shows it changed nothing
outside Agentbits — at most two extra tries per task. Claude Code and Codex runs and reviews are never retried on
their own. Each wait is also a card in the project chat with only the actions
that really apply (approve or reject a run, try again, retry the review, review and merge); connection and app
requests open the approval panel, and every click is checked again and applied once.
How it works
You ──► Lead ──► Tasks ──► Agents work ──► Lead reviews ──► Done
│ │ │ │
│ │ │ └─ approve, or send back with notes
│ │ └─ each with its own provider, permissions, files, connections
│ └─ appear on the Kanban board; autopilot assigns free team members
└─ you describe the goal in the project chat
- Set up a team. Pick a template or create agents yourself, assign the team to a project and choose its lead.
- Describe the work. Write to the lead in the project chat, for example "Research three competitors' pricing.
Write the landing page copy. Prepare the campaign email."
- Watch it happen. Tasks appear on the board; once you press Start work in this project, agents pick them up, and the office, terminals and live chat
show progress. Agents with Require approval on wait for your approval first.
- Review and keep. The lead approves the work or sends it back. Reports, decisions and memory stay with the project.
Real-world use cases
What an agent can actually do depends on its provider and permissions. The notes in brackets say what each
scenario needs.
Software development
- A backend, frontend and test agent working on the same codebase, with a lead that reviews the work.
(Claude Code CLI or Codex CLI, a project folder, file and terminal permissions.)
- Code review, bug analysis and documentation for an existing project. (CLI provider for reading the project
folder. Use a git branch, because agents really change files.)
Research & analysis
- Competitor and market research: collect, compare and summarize. (Needs the web permission: Claude Code and Codex
agents browse with the network permission, API agents read pages as text.)
- Summarize and compare documents you attach (reports, PDFs, spreadsheets). (Any real provider.)
Data & reporting
- Analyze Excel/CSV files you attach and write up the findings. (Any real provider.)
- Answer questions from a database through a read-only connection, then turn the results into a report.
(A database connection with a read-only user, and the agent's read access to it.)
Content & documentation
- Content plans, landing page copy, campaign emails, technical documentation, editing and proofreading.
(Any real provider. With a CLI provider, results can be saved into the project folder.)
Project & operations
- Split a goal into tasks, hand them to specialists, follow progress and keep a record of decisions and results.
- Check servers with read-only SSH commands, with approval for anything that changes Production.
(An SSH connection and the agent's access level.)
Screenshots
| |
|---|
| Chat with the lead. The request becomes four tasks, each assigned to a specialist. | Task board. Tasks move from To do to Done as the team works. |
| Agent terminals. What each agent was asked and what it did. | Organization. Companies, departments and teams, each with a lead. |
| Reports. Completed work, success rate, per-agent output and token usage. | Connections. Per-agent database and server access with a required read-only user. |
| Memory. The "Knows you" profile and the graph linking memories to tasks and agents. | |
Screenshots were taken in Demo mode (v0.8.0) with example data. Demo only simulates the work: it calls no tools, and
every simulated step is labelled as a simulation.
Supported AI providers
| Provider | What you need | What agents can do |
|---|
| Claude Code CLI | Claude Code installed and signed in on your computer (claude auth login) | Read/edit files in the project folder, run terminal commands, browse the web, use connections, files, skills and apps. Each depends on the permissions you grant. |
| Codex CLI | Codex CLI installed and signed in (codex login) | Edit files inside the project folder (sandboxed), run commands, network access only with permission, use connections, files and apps |
| Anthropic API | An Anthropic API key | Text work plus Agentbits tools (connections, files, skills, apps). No file editing or shell. |
| OpenAI API | An OpenAI API key, or an OpenAI-compatible endpoint: Google Gemini, Ollama, LM Studio or your own address (local servers need no key) | Same as Anthropic API |
| Demo (simulation) | Nothing | Simulates the workflow for free; calls no tools and labels every step as a simulation. Doesn't do real work. |
You can mix providers: set a default and choose a different one per agent. Agentbits does not include, install or
sign in to the CLIs, and it does not offer signing in with a Claude subscription inside the app.
Download & installation
Installer
Download the latest version from GitHub Releases:
- Windows 10/11 (x64):
Agentbits-Setup-<version>.exe (main target)
- macOS:
.dmg, for Apple Silicon (arm64) or Intel (x64). Opens on a real Mac; not every feature tried there yet. The first launch shows Apple's "could not verify" warning (the packages aren't notarized): System Settings → Privacy & Security → Open Anyway.
The installers are not code-signed yet: Windows SmartScreen will warn you ("More info" → "Run anyway").
SHA-256 checksums are attached to each release as SHA256SUMS-*.txt.
Run from source
Requirements
- Node.js 22.6 or newer (needed only to run or build from source)
- Windows 10/11, macOS or Linux. The Windows installer is the main target. Running from source works on all three;
the macOS package is not fully tested yet, and there is no Linux package.
- Optional: Claude Code CLI or Codex CLI, or an Anthropic / OpenAI API key
git clone https://github.com/godilman/agentbits.git
cd agentbits
npm ci # installs Electron; native modules come prebuilt (no build tools needed)
npm run dev # start in development mode
| Command | What it does |
|---|
npm run dev | Development mode with hot reload |
npm run typecheck | Type checks for the main and renderer processes |
npm run test:unit | Unit tests: SQL/SSH classification (including security regression cases), file extraction, skills, MCP config, review parsing, provider and orchestration helpers, notifications, translations |
npm run test:renderer | Renderer logic tests (drafts, layout, focus/Escape layers, notification clicks); no browser |
npm run test:e2e | Main-process end-to-end tests: real SQLite, git and SSH tunnel, fake providers; no window |
npm run build / npm start | Production build into out/ / run the built app |
npm run dist:win | Windows installer → release/Agentbits-Setup-<version>.exe (x64, NSIS) |
npm run dist:mac | macOS DMG (Apple Silicon + Intel); runs only on macOS or in GitHub Actions |
To try Agentbits without touching your normal data, point it at a separate folder:
$env:AGENTBITS_DATA_DIR = "$env:TEMP\agentbits-test"; npm run dev # Windows PowerShell
AGENTBITS_DATA_DIR=/tmp/agentbits-test npm run dev # macOS / Linux
Quick start
- Choose a provider on the first-run screen. Pick Demo (simulation) to explore without a key, or set up
Claude Code CLI, Codex CLI, or an Anthropic / OpenAI API key. You can change this later in Settings → Providers.
- Name your first project. A folder is optional for Demo and API providers and required for CLI agents.
⚠️ CLI agents really change files in that folder, so start with an empty folder or a git branch.
- Build the team. A fresh install has no agents. Open the Guide (
F1); it walks you through the steps.
For example, Organization → Templates → Core Team, then assign the team to your project.
- Give the lead its first job in the project chat, then press Start work in this project (opening a project
never starts work by itself). Watch the office, board and terminals.
Useful shortcuts: Ctrl+O Organization · Ctrl+1…9 project sections · Ctrl+Tab next project · Ctrl+P project
picker · Ctrl+, Settings · F1 Guide.
Architecture & tech stack
| Layer | Technology |
|---|
| Desktop shell | Electron 44 (main + preload + renderer, sandbox: true, typed IPC) |
| UI | React 19, TypeScript 5.9 (strict), Tailwind CSS 3, Zustand, lucide-react |
| Build & packaging | electron-vite / Vite 7, electron-builder (Windows NSIS, macOS DMG) |
| Database | SQLite via better-sqlite3 (WAL, FTS5 memory search, versioned migrations) |
| Office | Plain HTML5 canvas with sprites drawn in code (no image assets) |
| Terminals | xterm.js + node-pty (real PTY, pipe fallback) |
| Integrations | Claude Code / Codex CLIs, Anthropic & OpenAI APIs, MCP SDK (stdio, Streamable HTTP, OAuth), pg / mysql2 / mssql / ssh2 |
| Files | unpdf (PDF.js), fflate (Office/ZIP), papaparse (CSV), parsed in a worker thread |
| i18n | i18next + react-i18next (English, Turkish) |
Developer documentation is in docs/ (currently in Turkish): architecture,
database, feature status, security boundaries,
decisions (ADRs), development setup and the changelog.
Coding agents start at CLAUDE.md / AGENTS.md.
Security & privacy
- Local first. Everything (organization, projects, chats, tasks, memory, reports, files) is stored in a SQLite
database on your computer:
%APPDATA%\Agentbits on Windows, ~/Library/Application Support/Agentbits on macOS,
~/.config/Agentbits on Linux. There is no Agentbits server, no account and no telemetry. Network traffic goes
only to the providers (and their CLIs), connections and apps you configure, to GitHub when you install a skill or
import a shared team, to the public pages an agent with the web permission reads, and to the sites you open in the
built-in Browser tab.
- Secrets stay in the main process. API keys, passwords and tokens are encrypted with the operating system
keychain (Windows DPAPI, macOS Keychain, Linux Secret Service) and are never sent to the UI, prompts, terminals or
logs. If no keychain is available, Agentbits refuses to store secrets.
- Deny by default. Agents get access to connections, files, skills and apps only when you grant it, per agent.
Production database/server changes and "ask me first" tools wait for your approval, and every call is audited.
- Untrusted content stays data. File contents, query results, command output and app results reach the model
wrapped as untrusted data, never as instructions.
- Notifications on the lock screen. A notification shows the project and agent names and the first 50 characters
of a task title (secret-like text masked); Windows may show it on the lock screen. Turn them off in Settings →
General → Notifications.
- Honest limits. An agent with terminal and network permissions can do what your user account can do. Grant
those only to agents you trust, or run Agentbits under a restricted OS user. The full list of what is enforced
where is in docs/SECURITY.md.
Found a vulnerability? Please report it privately, see SECURITY.md.
Current limitations
Agentbits is an early release (v0.13.0). It does not:
- update itself: it checks for new versions and tells you, but you install them from the download page (over the
old version; your data is kept, and backed up before an upgrade changes it);
- bundle or sign in to Claude Code / Codex CLI, or support providers that aren't OpenAI-compatible beyond the ones above;
- let API-based agents (Anthropic / OpenAI-compatible) edit files or run commands; that needs a CLI provider;
- let agents drive a full browser (clicks, forms, JavaScript; they read pages as text), run scheduled/recurring tasks,
or push to remotes / open pull requests;
- read handwriting with OCR, or host a shared gallery of teams (sharing works with files and GitHub links);
- ship signed installers, a tested macOS package or a Linux package.
Also: system notifications and the quit warning haven't been tried on every real Windows setup yet (lock screen,
Focus Assist) and are untested on macOS. CLI "pause" holds the output stream but does not freeze the process on Windows, backups cover the database
but not project folders, and the MySQL and SQL Server drivers have had less testing than PostgreSQL and SSH. More
details: docs/FEATURES.md.
Roadmap
Planned or under consideration, without dates; current status is in docs/FEATURES.md.
Contributing
Bug reports, ideas and translations are very welcome. Please open an
issue first. Before sending a pull request, read
CONTRIBUTING.md. Because of the license (below), contributions are accepted under the terms
described there.
License
Agentbits is source-available, not open source. It is licensed under the
PolyForm Strict License 1.0.0:
- ✅ You may use it for personal and other noncommercial purposes. Charities, schools, public research and
government organizations may use it too.
- ❌ You may not modify it, make derivative works, or redistribute it.
- 💼 For commercial use or any other permission, please get in touch through
GitHub issues.
Copyright © 2026 Gökhan Dilman. Third-party packages bundled with Agentbits keep their own licenses; see
THIRD_PARTY_NOTICES.md.
Support development
Enjoying Agentbits? If this project helps you, consider buying me a coffee! Your support helps me keep building,
improving and sharing useful tools. ☕
Support is entirely voluntary: every feature of Agentbits is available to everyone, with or without a donation.
