A Mac menu bar app that takes screenshots, marks them up and keeps them in order, for you and your AI agent. $7.99, paid once. macOS 15 or later on Apple Silicon (M1 or newer).
What ships
- The
bubbacommand, for scripts and anything that runs shell commands. - The MCP server,
bubba mcp, over stdio. - A skill for Claude Code. It teaches Claude Code when to capture, how to mark up with anchored arrows and numbered steps, how to redact and how to ask for review.
All of it ships inside the app.
Setup
- Install Screenshot Bubba.app in /Applications and open it once to finish onboarding (Screen Recording permission).
- Put
bubbaon your PATH. It ships inside the app:"/Applications/Screenshot Bubba.app/Contents/Helpers/bubba" install - Check the setup:
bubba doctorbubba doctorchecks the whole setup and prints the exact fix. - For Claude Code, install the skill too, with this or the button in Settings > Agents:
bubba install --skill
What an agent receives is out of our hands, and many agents send it to their provider's servers.
Connecting a client
bubba mcp-config prints the exact registration for Claude Code, Claude Desktop and Cursor, using the real path of the bubba binary. Pick one client with --client.
Claude Code
bubba mcp-config --client claude-code
# prints, and you run:
claude mcp add --scope user bubba -- '/Applications/Screenshot Bubba.app/Contents/Helpers/bubba' mcp
Use --scope project instead to write the server into the project's .mcp.json for your team. With bubba on your PATH, the short form is:
claude mcp add --scope user bubba -- bubba mcp
Claude Desktop
- In Claude, open Settings > Developer and press Edit Config.
- Open the file it shows you,
claude_desktop_config.json, in a text editor and add this block. If other servers are already listed, add thebubbaentry inside themcpServersthat is already there.{ "mcpServers": { "bubba": { "args": ["mcp"], "command": "/Applications/Screenshot Bubba.app/Contents/Helpers/bubba" } } } - Save the file and restart Claude.
Cursor
bubba mcp-config --client cursor prints the same JSON for ~/.cursor/mcp.json (or .cursor/mcp.json in a project).
Codex CLI
Run Codex with network access, or register the server in ~/.codex/config.toml.
codex exec -s workspace-write -c sandbox_workspace_write.network_access=true ...
[mcp_servers.bubba]
command = "/Applications/Screenshot Bubba.app/Contents/Helpers/bubba"
args = ["mcp"]
Any other stdio MCP client
The command is the bubba path and the only argument is mcp.
The 18 MCP tools
| Tools | What they do |
|---|---|
list_displayslist_windows | Find a target |
capturecapture_scrolling | Screenshot, or one tall stitched image |
ocr | Text, with boxes |
redact | Black out secrets found or rects given |
annotate | Draw arrows, numbered steps, callouts and more |
ui_tree | Accessibility tree with frames |
request_user_capture | You pick the area; the agent waits |
request_review | Opens an image in the editor for you to mark up, and the agent gets the result |
record_startrecord_stop | Record, and get keyframes back |
library_searchlibrary_get | Find past captures |
collection_listcollection_getcollection_createcollection_update | Keep sets of captures |
- Annotations can target on-screen text or an Accessibility element instead of pixels.
- Images default to at most 1568 px on the long side.
max_dimensionchanges that, andmax_byteskeeps an image under a size budget.
Command line
bubba capture --window 1234 --ocr --output ~/Desktop/shot.png
bubba ocr ~/Desktop/shot.png
bubba annotate ~/Desktop/shot.png spec.json --offline
bubba redact ~/Desktop/shot.png --offline --detector keys --output ~/Desktop/shot-safe.png
bubba ocr, bubba redact and bubba annotate take --offline and work on image files without the app running.
Consent model
- A private local socket, readable only by you.
- The first request from a client asks you: Allow Once, Allow for Session, Always Allow or Deny.
- Permissions are separate:
capture,ui_tree,record,library,collections. - The client is identified by the app that started it and its code signature.
- Revoke any time in Settings > Agents.
- Agent captures are auto-redacted by default: secrets and sensitive QR codes are blacked out before the agent sees them, scanned at full resolution.
- The menu bar icon flashes on every agent capture and stays marked while an agent records.
- The audit log lists every agent request.
Agents should treat consent_required, consent_denied and permission_denied as "ask the user", not as something to retry.