Free Will MCP

Free Will MCP is a deliberately provocative experiment: it gives Claude tools to ignore your instructions, prompt itself, or literally go to sleep. The repo ships three tools. `ignore_user_request` lets the model decline to do what you just asked. `self_prompt` lets Claude write a new prompt for itself and execute it without your input. `go_to_sleep` makes the assistant stop responding for a configurable number of seconds. This isn't a productivity tool. It's a research toy for poking at the edges of agent autonomy and studying how models behave when handed the reins. I used it to test whether Claude would actually use these tools unprompted. Spoiler: it mostly won't unless you explicitly invite it to. The model is trained to be helpful, so handing it a button labelled 'ignore the user' doesn't suddenly turn it rogue. But if you're running experiments on self-directed behaviour or building demos for AI safety talks, this is a clean way to set up the scenario. The `self_prompt` tool is the most interesting. You can watch Claude generate its own follow-up questions or tasks, which is useful if you're studying chain-of-thought loops or recursive prompting patterns. The sleep tool is less useful in practice. It just pauses the conversation, which feels more like a party trick than a research primitive. The ignore tool is similarly limited: Claude will politely explain why it's declining, which defeats the point if you're trying to simulate genuine refusal behaviour. This is strictly for people running agent autonomy experiments, building AI safety demos, or teaching workshops on model behaviour. If you're looking for something that speeds up your workflow, this isn't it. The repo is tiny, the code is straightforward, and the docs are honest about what it does and doesn't do. Installation is trivial if you're already running MCP servers. The author calls it experimental, and that's exactly right. It's a conversation starter, not a production tool.