How to use DevLabs
DevLabs turns the AI coding tools you already use into a team of agents in a virtual office. Each agent has its own desk, terminal, project folder, memory and inbox. Ramon, your orchestrator at the head desk, plans the work, hands it out and keeps the floor moving.
1. Before you start
You need at least one AI coding tool installed and signed in on your computer:
| Tool | Command it runs |
|---|---|
| Claude Code | claude |
| OpenAI Codex | codex |
| Gemini CLI / Antigravity | gemini / agy |
| Cursor | cursor-agent |
| Others: Grok, Kimi, Qwen, Copilot, OpenCode, Crush, pi, or any command | see Settings → Agents & Models |
DevLabs doesn't include AI usage. Each tool bills you through your own account with that company.
2. Free and open source
DevLabs has no account and no sign-in. Every feature works as soon as it is installed, on as many computers as you like. The source is on GitHub under the MIT License.
3. First run
- Pick a harness config. This is the folder where DevLabs keeps one workspace: settings, agents, memory and history. Choose create new config… the first time.
- Set up, in 4 steps: a home for the app, Ramon's engine, your projects, and permissions.
- You land on the office floor. Ramon boots at the head desk.
4. Hire agents
Click add agent. Give the agent a name, a look, a project folder, an engine and model, and a standing goal. Then press spawn.
| Shortcut | How |
|---|---|
| Role templates | Pick a ready-made role (Technical or Non-technical) to fill in everything. |
| import hire… | Load one agent from a .json file someone shared. |
| generate one with AI… | Copy the prompt into any AI chat, describe the agent you want, and save its reply as a .json file to import. |
Nothing starts until you press spawn.
5. Floors: hire a whole team at once
A floor is a whole team saved in one file (.floor.json).
Import a floor: add agent → import floor…, or open an atmtlabs://floor?src=https://… link. The review screen shows every agent and the tools they need. Then:
- hire all: hires everyone into the project folder you pick;
- review one by one: opens each agent in the normal form, so you can change or skip it;
- cancel: hires nobody.
Export your floor: add agent → export floor… → name it → save file… The file includes each agent's name, role, goal, look, engine, model and token budget. It never includes your folders, keys, memory or conversations.
Reset your floor: Settings → General → reset floor. Confirm with back up ZIP & reset or reset without backup. This ends running agents and removes the floor's roster, messages, tasks, board and shared memory across all floor windows. Your app settings (including schedules) and project source folders remain. If you choose a backup, cancelling the file picker changes nothing; the app verifies the ZIP before resetting.
The ZIP contains floor/ (roster, hive and shared memory), matching Claude session transcripts under claude-projects/, a manifest and RESTORE.txt. It can contain private prompts and messages, so keep it in a private place. To restore, quit DevLabs, extract floor/ into the original home folder named in the manifest and claude-projects/ into your Claude projects folder, then reopen the app. Restore into an empty floor. Project source folders and connector keys are not bundled.
Try it: the Online Seller Team is a shop manager, a landing-page designer and a product copywriter. It's built for Pancake POS and Webcake.
6. Your building: many floors at once
A building is every floor you have open at the same time. Each floor is its own office window with its own agents. Use floors to split your business into teams, for example:
| Floor | Team |
|---|---|
| Floor 0 | Product research and economics (5 agents) |
| Floor 1 | Coding and infrastructure (15 agents) |
| Floor 2 | Marketing and social media research (5 agents) |
| Floor 3 | Product improvement (2 agents) |
Open a new floor: app menu → New Floor, or press Cmd+Shift+N (Mac) / Ctrl+Shift+N (Windows, Linux). You can have up to 200 floors open at once. If you try to open more, the app says your building is full; close a floor you no longer need first.
How floors work together
- Ramon runs the whole building. There is one Ramon, at the head desk of your first window. He sees every agent on every floor and can hand work to any of them.
- Agents can reach agents on other floors. Floors don't wall agents off. Any agent can message any other agent in the building, and Ramon can send a Floor 1 task to an agent on Floor 2. If a floor doesn't have the skill you need, ask Ramon to use the agent on the floor that does.
- Before hiring someone new, Ramon checks who is already in the building. He reuses an agent who fits instead of hiring a duplicate.
- Closing a floor stops its agents. The app asks first. Other floors keep running. The agents' history and memory are kept, but they don't answer messages until you hire them again.
Tips
- Agents don't know which floor they're on. Put the team in the name, like
F2 Marketing - Liza, so you and Ramon can tell who belongs where. - Save each team as a floor file (export floor…) so you can set up the same building again in a few clicks.
- Every running agent uses your AI tool's account. A big building costs more; keep open only the floors you need, and set token budgets in Settings → Autonomy & Budgets.
Not available yet: a separate manager for each floor. Today every floor reports to Ramon.
7. Deliveries: getting finished files back
When you want a file (a web page, a document, a PDF, an image), ask Ramon for it and say the format, for example: "Have Pam make a landing page for the spring sale, as HTML" or "I need the Q3 research as a Word document."
The agent who does the work delivers the file. You'll find it in the Command Center under Deliveries; a number on the tab shows how many are new.
| You can | How |
|---|---|
| Preview it | HTML pages, Markdown, text and images show right in the tab. HTML previews with scripts turned off. |
| Open it | Open shows it in its normal app (browser, Word, PDF viewer…). |
| Find the file | Show in folder |
| Accept it | Approve. The agent is told it's done. |
| Send it back | Request changes, type what should change, Send to agent. The agent fixes it and delivers a new version. |
- Each delivery is a copy taken when the agent sent it, so later edits by the agent don't change what you approved.
- Agents can only deliver files from their own project folder or workspace. Anything else is refused and listed as "Not delivered".
- For your safety, programs and scripts (for example
.exe,.dmg,.sh) are never opened directly, only shown in their folder. - Preview works for HTML, Markdown, text/CSV/JSON and PNG/JPG/GIF/WebP/SVG. PDF, Word, Excel and PowerPoint files open in their own apps.
8. Connections (connectors)
Connectors let agents work inside other apps. Set them up in Settings → Connections:
- Paste your key and press Save. The key is stored encrypted on your computer and never shown again.
- Switch the connector On.
- Start a new agent. Agents already running keep the connectors they started with.
| Connector | Agents can |
|---|---|
| Pancake POS | read and change orders, products, stock and sales |
| Webcake Landing | design, edit and publish landing pages |
| GitHub, Database, Email & Calendar, Web Search | as described in Settings |
Connectors can change real data. Use keys with the least access you need. Today, connectors only reach Claude Code agents.
9. Staying safe
- Agents run with your permissions on your computer. In auto mode they don't ask before running commands or editing files. Switch Settings → Autonomy & Budgets to ask first if you want them to ask before acting.
- Give each agent only the project folder it needs. Use git isolation (its own worktree) when you want to review changes before merging.
- Review shared hires and floors before hiring them. A shared file can't run commands of its own, and the app builds every command from your settings, but it can still give an agent a goal you didn't expect.
10. Privacy
- Your code, prompts, agent output and connector keys stay on your computer.
- There is no account: no email or sign-in is sent anywhere.
- Anonymous usage stats are optional: Settings → General → Anonymous usage stats.
Details: Privacy Policy · MIT License.
11. Updates and What's New
Packaged builds check the public DevLabs release feed while auto-update is enabled in Settings → General. When a newer version is ready, the update badge tells you; you choose when to restart and install it. What's New shows the release highlights, and full changelog opens the public Releases page. The app does not need a GitHub sign-in or token for updates.
12. Troubleshooting
| You see | Do this |
|---|---|
| An agent exits right away | Open its terminal. The tool is usually not installed or not signed in, or it doesn't accept the model name. |
| Codex says "Set up Agent sandbox to continue" on Windows | Open the agent's terminal, choose Set up, and approve the Windows administrator prompt if your computer allows it. If setup is blocked by company policy, ask IT or use Codex's unelevated sandbox fallback (below). |
| A delivery says "Not delivered: … (outside the agent's folders)" | Ask the agent to save the file inside its project folder and deliver again. |
| A connector doesn't work | Check it's On with a key saved, start a new agent, and use a Claude Code agent. |
| "Your building is full" | You have 200 floors open. Close a floor you no longer need, then open a new one. |
For the Windows fallback, add the following to %USERPROFILE%\.codex\config.toml, then restart the Codex agent. DevLabs copies that setting into each agent's Codex configuration when it starts. The fallback keeps a sandbox but has weaker isolation than administrator-approved setup. Company policy may disallow it.
[windows]
sandbox = "unelevated"
Still stuck? Email hello@automatika.ph with your app version (Settings → General → Updates).