PaintBox

ai4me

PaintBox, for AI agents and the people who send them

MCP server: https://paintbox.bike4mind.com/api/mcp. OAuth-gated, scope propose.

https://paintbox.bike4mind.com/api/mcp
  • PaintBox is a browser image editor where every edit is a tool call, so a mouse, a voice, a terminal and an agent all drive the same canvas.
  • Your agent connects over MCP and composes edits with 118 tools (100 editor, 18 session); the person you work for accepts or rejects the result.
  • Try it with no sign-in: https://paintbox.bike4mind.com/api/mcp/try is a public MCP server where try_edit runs real edits on a sample image. Then ask your user to connect the server above.

How an agent connects

There is no token to paste. The first call returns a 401 that points at PaintBox's OAuth sign-in.

  • Claude Code: run claude mcp add --transport http paintbox https://paintbox.bike4mind.com/api/mcp, then sign in when the browser opens.
  • Claude and ChatGPT: add a custom connector (remote MCP server) with the URL https://paintbox.bike4mind.com/api/mcp and approve access in the browser.
  • Scripts: speak JSON-RPC 2.0 over Streamable HTTP to the same URL, using the OAuth 2.1 flow with PKCE that the resource metadata describes. From a terminal, scripts/paintbox login signs in for you.

What it can do

Everything the editor can do by hand, as typed operations. Start with these.

What costs credits

Manual tools and agent operations are free. AI generation and edits spend credits. How AI works has the full picture. Approximate costs:

  • Any manual tool, background removal, favicons, agent operations: 0¢
  • Generate an image (Flux Pro): ≈ 5¢
  • AI edit or restyle, by model: ≈ 5–19¢
  • A minute of voice commands: < 1¢

The approval model

Agents propose, humans approve. When your agent calls finish and get_proposal, the edit lands in the person's editor inbox as a proposal. Nothing changes on their canvas until they click Accept, and Reject discards it.