Connect an Agent
Connect Claude, Codex, Cursor or Pi to edit photos with existing subscription credits and return results to the ClearCrowds editor.
Let your usual Agent fix your photos using your existing ClearCrowds subscription and personal credits.
Open the editor and connect an Agent →Connect and install
- Sign in to ClearCrowds, choose Connect an Agent in the editor sidebar, and select your assistant.
- Codex can open with the setup prompt prefilled. In other local clients, copy and send the setup prompt. It includes both MCP configuration and personal image Skill installation. Cursor's native install link adds only MCP; send the full prompt to install the Skill too.
- Check your account in the browser, choose Allow connection, then return to the client to complete sign-in.
- The Agent checks your account and available capabilities. Setup and connection checks do not generate images or spend credits.
My connections shows verified connections. Use the pencil to name a device, such as “George MacBook Pro · Codex.” Existing connections can copy the Skill setup instructions to add the Skill. A working MCP connection does not prove local Skill installation; the client must verify the file and refresh discovery.
Invoke ClearCrowds
| Client | Explicit invocation | Skill location |
|---|---|---|
| Codex | $clearcrowds | ~/.agents/skills/clearcrowds/SKILL.md |
| Claude Code | /clearcrowds | ~/.claude/skills/clearcrowds/SKILL.md |
| Cursor | /clearcrowds | ~/.cursor/skills/clearcrowds/SKILL.md |
| Pi | /skill:clearcrowds | ~/.pi/agent/skills/clearcrowds/SKILL.md |
You can also ask “Use ClearCrowds to fix this photo” in clients that select matching Skills automatically. If a new Skill is missing, refresh or start a new session as required by your client. Pi supports /reload.
Browser-based Claude and Grok use remote MCP and natural language; connecting does not install a Skill on your computer. Other clients choose their supported setup method. Pi needs an installed adapter that supports remote MCP; there is no universal Pi MCP install command. There is also no universal @clearcrowds command across clients.
Try editing a photo
- Choose an assistant in My connections, then Share a photo and explicitly select an image.
- Try a city background, text replacement, or background people removal. Edit the request and send the resulting prompt. It identifies the exact selected photo.
- The Agent directly edits the photo using existing credits after your clear request, without asking for credit-cost confirmation. It returns a preview and an editor link; the result is saved in that photo's edit history.
For example:
Use ClearCrowds to replace the background of my shared photo with a city at night. Preserve the person's appearance and pose, and blend the lighting naturally.You can also select a local image for the Agent. The client uploads it before editing. A remote MCP server cannot read a local path or chat attachment by itself; attachment access depends on the client.
Manual MCP configuration
Remote URL: https://app.clearcrowds.com/api/agent/mcp. Transport is Streamable HTTP with browser OAuth. Do not paste passwords or tokens into chat.
These commands target the production service. For local testing, use the actual address generated by the editor.
Codex
codex mcp add clearcrowds --url https://app.clearcrowds.com/api/agent/mcp
codex mcp login clearcrowdsClaude Code
claude mcp add --transport http --scope user clearcrowds https://app.clearcrowds.com/api/agent/mcp
claude mcp login clearcrowdsYou can also authenticate through /mcp in Claude Code. In Cursor, Claude and other remote MCP clients, add the remote address and complete browser authorization; settings locations depend on the client version.
Download the personal image Skill (SKILL.md) and save it in the corresponding directory above. It includes the MCP address, explicit photo selection, local uploads, text-to-image generation, direct editing, job polling and returning to the editor.
Generate an image from text
Tell your connected Agent what you want, for example: “Create a landscape image of a small cabin beside a lake at sunrise.” No upload or shared photo is needed.
The Agent calls quote_generation with your prompt and optional aspect_ratio (1:1, 3:2, 2:3, 16:9, 9:16), then immediately calls generate_image with that quote. A clear generation request authorizes use of existing personal credits; no credit-cost confirmation step is required. An explicit budget still applies. Capacity is reserved at submission and charged only after the result is durably delivered. Failed tasks release the reservation.
The Agent reuses the exact quote on a retry, polls get_job, and returns the preview and editor_url. Results appear in your ClearCrowds history and can be edited through the same connection. Setup checks alone never generate images or consume credits.
Credits, permissions and troubleshooting
- Which balance is used? The same subscription and personal balance as the editor. No developer project, API key or API Credits top-up is needed. Insufficient credits return a management link; credits are never purchased automatically.
- Missing photo? Explicitly share it with this connection, then send a fresh editing prompt.
- Signed in but tools are unavailable? Refresh the client's MCP tools and check
accountandcapabilities. Opening an authorization page alone does not complete the token exchange or verify the connection. - Why a system credential prompt? The client manages OAuth credential storage. ClearCrowds does not read your keychain or ask for credentials in chat.
- How do I disconnect? Disconnect in My connections to revoke access immediately. The Agent can access only explicitly shared photos, images uploaded through its connection and results created through that connection, not the entire library.
Developer integrations use separate API Credits and CLI workflows.
