Muse Code: what our integration actually verifies

Read the evidence before the compatibility claim

The repository records Muse Code 1.3.0 discovering the published MCP tools and executing a Stop hook. The record uses an echo provider. It does not verify a model-driven app-building session, pre/post hooks or observer behavior. That distinction is the starting point for trying the integration.

Install outside the agent sandbox

The project setup guide uses a global install, followed by a settings entry:

npm install -g ios-agent-mcp@latest

Follow the Muse setup section and merge its configuration with your existing settings. Preserve the schema version and other servers. If the command cannot be found, inspect the installed executable path instead of guessing a new location.

Verify in stages

1. Confirm the server starts and tools are discoverable.
2. Ask for a read-only review of a small synthetic Swift file.
3. Inspect the returned finding and its source location.
4. Only then try a controlled edit and an available build/test step.

Record the Muse version and npm package version alongside the outcome. A successful hook invocation does not prove that the agent understood the failure or repaired it correctly.

Keep failures useful

If a step fails, distinguish an unavailable executable, an invalid configuration, a tool error and a model decision. Keep private project source out of public reports. A small synthetic reproduction is easier to inspect and safer to share.

This article describes the project's recorded test scope, not a newly verified Muse release or a claim that its background observer works with every subagent. Future compatibility claims should be backed by a new reproducible record.

Inspect the exact verification record

Decision checkpoints