The value is entirely in when it gets called: with it wired in, a long agent run stops and shows its work instead of finishing on an assumption, and you steer with a sentence rather than reviewing afterwards. Images in replies make the correction cheaper — a screenshot of the wrong output says more than a paragraph describing it. The stored prompt templates are worth setting up early if you keep giving the same correction.
A feedback checkpoint. The assistant hands over a summary of what it just did, a page opens in the browser, and it waits there for your direction instead of ending the turn on its own guess.
- One call that reports the work done and blocks until you answer in the browser
- Replies that carry images as well as text, with uploads compressed and thumbnailed on the way in
- Reusable prompts: templates stored, categorised and searchable, with dynamic substitution and version history, so a direction you give often does not get retyped
- Preset scenes for recurring kinds of feedback, importable and exportable
- Sessions kept in a local SQLite store with expiry, so a reply is not lost when a window is closed
- The web port allocated automatically, with conflict detection, rather than fixed
Nothing to authorize; it runs as npx -y smd-mcp-feedback-collector@latest and opens a page in the browser on the machine where it runs, so the person answering has to be at that machine. Older releases take MCP_TRANSPORT_MODE and MCP_WEB_PORT explicitly, and there is a separate service for intercepting prompts through the Stagewise toolbar. The tool's own instruction is that the assistant should call it whenever it would otherwise summarise a task or ask for an opinion — that convention is what makes it fire at the right moment.
One command — npx -y smd-mcp-feedback-collector