MCP for agents

Use the desktop MCP bridge when an agent should create and inspect assets inside a local SorryAssets project.

Connect the desktop

Start the installed app, then register its bundled bridge once with your agent host. On macOS the bridge is located at:

/Applications/SorryAssets.app/Contents/Resources/mcp-bridge/bridge.js

Codex configuration:

codex mcp add sorryassets -- node \
  /Applications/SorryAssets.app/Contents/Resources/mcp-bridge/bridge.js

Claude Code configuration:

claude mcp add --scope user sorryassets -- node \
  /Applications/SorryAssets.app/Contents/Resources/mcp-bridge/bridge.js

This registration is local setup and does not require a SorryAssets account key. The bridge reads the running app's current loopback address and launch token. Do not copy either token or a provider credential into host configuration.

After registration, start SorryAssets before asking the agent to work. The app does not need to stay in the foreground.

What the agent can do

Desktop MCP can create a project in the configured workspace, list and open registered projects, inspect and edit the graph, import local files, discover the public catalog, estimate and start generation, observe delivery, and export results.

The tools do not accept arbitrary project paths or scan the disk. Generation returns before long-running work is complete, so the agent reads the project graph until the task is terminal and the output has a local file path.

Reasoning, Skill instructions, and permitted local Skill scripts run in the agent host. SorryAssets provides atomic tools and durable project state. It does not execute a Skill workflow.

Direct HTTP

Developer integrations that own their output storage can use the public HTTP API. Direct HTTP does not create desktop project lineage. See Direct HTTP API.

Security boundary

Agents receive public model and task facts. SorryAssets never returns provider credentials, route or connection ids, wholesale prices, raw upstream model ids, or upstream asset URLs.