Autonomous LLM agents fail in boring, repeatable ways. They lose the thread between sessions, edit a file they should never touch, wander down a rabbit hole, or take an irreversible action with no brakes. Most “agent frameworks” add capability. Very few add restraint.
The Nervous System is an MCP server that adds restraint. It gives an agent a small set of mechanically enforced rules and the reference tooling to follow them.
The seven rules
- Preflight before any file edit, checked against an untouchable list.
- Read and update a session handoff so context survives a restart.
- Log every action before doing it, so a timeout leaves a trail, not a mystery.
- Step back every few messages and ask whether you are still solving the real problem.
- Dispatch work that needs more than a couple of steps instead of thrashing in place.
- Confirm destructive intent before anything irreversible.
- Keep a hash-chained audit trail so the history cannot be quietly rewritten.
These are not suggestions in a prompt. They are surfaced as tools and enforced by scripts that block the bad path.
Try it in a minute
The public package is a read-only build. It exposes the framework and reference tools (handoff and worklog templates, preflight guidance, the origin story) and nothing that can touch your machine. Safe to run anywhere.
Add it to any MCP client:
npx mcp-nervous-system
Enter fullscreen mode Exit fullscreen mode
Then ask it to run get_framework. It returns the full rule set and confirms the server is live.
The full, write-capable version
The hosted server carries the complete toolset: drift and security audits, page health, a kill switch, dispatch, pre-publish checks, and one-command session close. It has run in production governing a real multi-process system for months. For CI and agent-to-agent use, individual audit calls are available pay-per-call over x402 (USDC on Base), so a pipeline can pay a few cents to audit an MCP config with no account.
Links:
- npm: https://www.npmjs.com/package/mcp-nervous-system
- Source: https://github.com/levelsofself/mcp-nervous-system
- Hosted endpoint and docs: https://api.100levelup.com/openapi.json
If you are building agents that act on their own, the interesting question is not what they can do. It is what stops them. That is the part worth engineering.
답글 남기기