Sweepspace guide

1

Install and first start

Sweepspace runs on Windows 10 and 11 (64-bit). Download the installer from the Pro page (the beta, free for now) or the home page (the free version).

SmartScreen may warn you. The installer isn't code-signed yet, so Windows doesn't know it. Choose More info, then Run anyway. To check you got the real file, compare its SHA-256 with the one on the download page: in PowerShell, run Get-FileHash .\Sweepspace_0.3.0-beta.5_x64-setup.exe.
Installing over an older version keeps your projects and settings. That includes 0.2.1 and earlier betas.
On first start Sweepspace fetches what's missing. Git and OpenCode come from their official sources, each checked against its published checksum. Progress shows at the bottom. Settings › Getting ready lists what's installed and what isn't.
Open a folder. Pick your project. Sweepspace builds a map of its files, imports and functions in a few seconds, without uploading anything.

No subscription? You can still use everything. Without Claude Code, bots, the CEO's hires and the crew run on OpenCode's free models.

2

Agents and the map

Start an agent opens one in a terminal beside the others. Add more agents › Install sets up Claude Code, Codex, Gemini CLI, Antigravity, Kimi or Aider. It fetches Node.js or uv first if your PC has neither.
Claude Code needs a sign-in once: type claude in any Sweepspace terminal and follow the login. If a pane says "Not logged in", that's the fix.
Every installed agent gets Sweepspace's map tools (overview, search, outline, read one function) on start, so it reads less to do the same job. Savings counts the tokens saved.
With two or more agents open, they see who is working on which files and message each other in Team chat under the terminals. You can write to all of them there too.

3

Board and the crew

Write a task on the Board, then hand it to the crew. Four roles take it from there:

Scout: Reads the code and writes the plan. Changes nothing.
Forge: Builds it on a branch of its own, so your work stays untouched.
Warden: Reviews the change in a fresh session and blocks anything risky.
Ledger: Keeps the budget and picks the model for each step. Not an AI.

You see each step live, can whisper to the bot at work, and merge the branch yourself when you're happy with it.

4

Bots

Bots are coworkers you talk to. Each one remembers the conversation, has a job, and can be given a browser, a shell and your project's files. Open Bots in the top bar, then Hire one from the catalog or Write your own.

Where it works

This PC

Works here like an agent in a terminal. Commands and edits in the project run straight away, and its browser is an Edge window of its own. It asks you before anything risky or outside the project. The default.

Its own computer

A Linux desktop in Docker, kept apart from your files. You can watch its screen live, take control and hand it back. Needs Docker Desktop.

Read only

Reads the project and the web, changes nothing. For research and answers.

What it may use

Browser: clicks, types and fills forms on real sites. It reads pages as text with named buttons, so it doesn't guess from screenshots.
Shell: runs commands. Turn it off and the bot has no shell at all.
Files: reads and edits the project.
Desktop: sees and uses its whole screen (its own computer only).
Your MCP servers: Claude Code bots can use the MCP servers from your Claude Code settings (Notion, Apify, a database…). Lookups run freely; anything that changes something asks you first.

Its brain

Pick Claude Code (Sonnet, Opus, Haiku) or OpenCode with a free model, in the form or under the message box. A free model that's busy or rate-limited is swapped for another free one automatically.

Each bot is a small Markdown file. Project bots live in .sweepspace/bots/ and travel with the code through git. Personal bots are yours in every project. Edit the file and the bot changes.

5

What a bot asks you about

When a bot wants to do one of these, a card appears with Allow once, Always allow and Deny. The card shows the whole command.

Always asks

Sending, posting, uploading, paying
Deleting folders, force-resetting git, pushing
Reading keys, passwords, .env and other secret files
Files outside the project
Changing its own rules, git hooks or agent settings
Web addresses stuffed with data
A tool from your MCP servers that changes something

Just does it

Reading the project and the web
Editing project files
Running tests, builds and installs
Browsing and reading pages
Lookups in your MCP servers

What it did, beside the conversation, lists every action and the rule that decided it. Budgets per run and per day stop a bot before it overspends, and near the end of a run it's told to wrap up instead of being cut off.

A bot on your PC reads the web, and a page can try to talk it into things. The asks above catch the common ways of sending your data out, but a shell that can run anything can still find a way round them. For a project full of secrets, use a bot with its own computer, or keep Shell off.

Bots on OpenCode can't stop and ask you mid-run, so a risky step is simply refused and the bot tells you what it wanted to do. Claude Code bots ask.

6

The hard things, step by step

A bot that does real work on your PC

Bots › Hire › Write your own. Name it, and write its role the way you'd brief a person: what it does, and what to say when it isn't sure.
Under Where it works, keep This PC. Tick Browser, Shell and Files.
Give it a job: "Check the coupon field on staging, fix it if it's broken, then tell the team."
The first time it uses the browser, Sweepspace sets it up (and fetches Node.js if your PC has none). That takes a minute once.
Answer its questions as they come. Always allow remembers the answer for that bot and that kind of command.

A bot that works in sites you're logged into

Give the bot the Browser tool and start any job that opens the site. Its own Edge window appears, with a profile only that bot uses.
Sign in yourself in that window. Bots never type passwords: password fields always come back to you.
The login stays in the bot's profile for next time. Your own Edge, its cookies and saved passwords are never touched.

A CEO that runs a team

Hire the CEO from the catalog.
Give it a whole goal: "Get the spring sale ready: coupon works, docs say so, and someone checks it on a phone."
It writes a plan, reuses the bots you have, hires up to three more, and hands each one a job. Their reports come back to it by themselves.
Anything that changes code goes to the crew, so you get a reviewed branch instead of a bot editing your project directly.
It ends with what was done, by whom, and what's left for you (usually the merge).

Bots with a computer of their own

Install Docker Desktop (free for personal use) and start it.
In the bot's form, choose Its own computer. The first one builds once: about 2.8 GB, a few minutes on a fast connection.
Watch it in Its screen. Take control lets you drive; Hand back returns it.
Each running computer needs about 700 MB of free memory, and two run at most. If Docker isn't running, the bot works on your PC that time and asks before every command.

Free models only

Without Claude Code installed, Sweepspace uses OpenCode's free models everywhere, with no account needed.
Free models get busy. A bot waits, retries, then moves to another free model by itself (up to four). If all are busy, Continue later picks the run up.
Keep jobs small and concrete: free models do much better with one clear task than with a whole project.

Your own rules

Rules decide what bots may do without asking. Yours live in %APPDATA%\ai.sweepspace.app\rules.yaml. A project can add .sweepspace/rules.yaml, but a project's file can only make bots stricter, so a cloned repo can't loosen anything. The strictest matching rule wins.

# let the CI Doctor run npm without asking
- when: { bot: ci-doctor, intent: shell, text: npm }
  do: allow
# never click "Submit" on any GitHub page
- when: { host: '*.github.com', intent: submit }
  do: block

Fields are bot, intent (read, activate, type, submit, file, shell), host and text. Passwords, payments and secret files always ask, whatever the rules say.

7

When something's off

"Windows protected your PC": That's SmartScreen and an unsigned installer. Choose More info, then Run anyway, after checking the SHA-256.
"Claude Code is signed out": Run claude in a Sweepspace terminal and sign in, then press Continue in the bot's conversation.
"Docker Desktop isn't running": Start Docker Desktop and wait for it to say it's running. Until then, bots work on your PC instead.
A free model is busy: The bot retries and switches model by itself. If it stops, wait a few minutes and press Continue, or pick another model under the message box.
The bot's browser didn't open: The first run needs the internet to fetch the browser tools (and Node.js if missing). Check Settings › Getting ready, then try again.
An agent doesn't use the map: Restart the agent from Sweepspace: the map tools are added when it starts. Agents you switched off in Savings stay off.
A bot keeps asking for the same thing: Answer Always allow, or add a rule (above). Safety asks (secrets, uploads, deleting, pushing) can't be turned off.
Sweepspace closed during a run: The bot says so when you come back. Press Continue and it picks up where it was.

8

Where things live

Project bots: <project>\.sweepspace\bots\
Team documents: <project>\.sweepspace\team\
Project rules: <project>\.sweepspace\rules.yaml
Your rules, bots, logs: %APPDATA%\ai.sweepspace.app\
Bots' browser profiles: %APPDATA%\ai.sweepspace.app\bots\browsers\
Tools it set up: %APPDATA%\ai.sweepspace.app\tools\

Found a bug or a gap in this guide? Open an issue on GitHub.

Running in five minutes. Then the hard things.