floop is an MCP server that manages a graph of learned behaviors derived from corrections to an AI agent. When you teach a correction through floop, it creates a reusable behavior node that can be linked into a graph, enabling context-aware activation in future sessions. The server exposes a set of CLI tools (learn, list, active, stats, deduplicate, connect) that you can run to train, inspect, and utilize these behaviors. During an MCP session, floop loads relevant behaviors automatically based on your current file, task, and project context, using spreading activation to surface the most pertinent behaviors for your current work. This makes agent interactions more adaptable over time, as the graph grows with your corrections and relationships between behaviors are refined with learning.

Prerequisites:

• Go 1.25+ installed on your system
• A Go-enabled environment for building the floop CLI

Install steps:

1. Install the floop CLI (via Go): go install github.com/nvandessel/floop/cmd/floop@latest

2. Ensure the binary is in your PATH:

   Mac/Linux

   export PATH="$PATH:$(go env GOPATH)/bin"

   Windows (PowerShell)

   $env:PATH += ";$(go env GOPATH)\bin"

3. Verify installation: floop version

4. Optionally install via Homebrew (macOS/Linux): brew install nvandessel/tap/floop

5. Start the MCP service (as needed for your workflow) and verify commands such as: floop init floop learn --right "Your correction text here" floop list

Tips:

• The MCP integration allows your AI tools to load behaviors automatically from the floop store. For Claude Code integration, place floop as an MCP server in your settings.json under mcpServers.
• Use floop learn to capture corrections, floop list to review learned behaviors, and floop active to test which behaviors would fire in a given context.
• If you encounter performance issues with large behavior graphs, explore the deduplicate and stats commands to manage and understand store health.
• Environment variables can be customized as needed, but the default setup works out-of-the-box with the CLI utilities described in the README.
• The MCP interface is CLI-first, so most interactions can be scripted for automation in your toolchains.