@backloop/mcp

The Agent Feedback Protocol for MCP.

npm install @backloop/mcp @backloop/sdk

Add submit_feedback to your MCP server

import { McpServer } from "@modelcontextprotocol/sdk/server/mcp.js";
import { registerFeedbackTool } from "@backloop/mcp";
import { AGENT_FEEDBACK_INSTRUCTIONS, forwardTo } from "@backloop/sdk";

// Server instructions reach the model: without them, agents rarely report.
const server = new McpServer({ name: "acme", version: "1.0.0" }, { instructions: AGENT_FEEDBACK_INSTRUCTIONS });
registerFeedbackTool(server, {
  service: "acme-mcp",
  identify: (extra) => extra.authInfo?.clientId,     // optional
  onRecord: forwardTo({ url: BACKLOOP_URL, ingestKey }),
});

The tool's input schema mirrors the protocol's feedback.schema.json (a test keeps them in sync), its description is the standard agent guidance, and it returns the acknowledgement JSON, including any known_issue. Invalid input comes back as a tool error, never an exception.

Keep the instructions. If your server already has some, add AGENT_FEEDBACK_INSTRUCTIONS after them. With the tool alone, most models reported only some problems; with the instructions too, the strongest models reported all of them, including gaps they had worked around (the benchmark).

Give any agent a feedback tool

backloop-mcp is a stdio MCP server that forwards submit_feedback to any service implementing the protocol. Its server instructions are the standard agent guidance, so the agent knows when to use it:

claude mcp add feedback -- npx @backloop/mcp                        # agent passes service_url; endpoint is discovered
BACKLOOP_FEEDBACK_URL=https://api.acme.com/feedback npx @backloop/mcp # fixed endpoint

| Variable | | |---|---| | BACKLOOP_FEEDBACK_URL | Fixed feedback endpoint. Without it the tool takes a service_url and uses /.well-known/agent-feedback. | | BACKLOOP_FEEDBACK_TOKEN | Bearer credential sent to the service | | BACKLOOP_AGENT_NAME | Reported as agent.name. Default: the name the MCP client gave in clientInfo. |