A new agent profile called toolsmith is reshaping how homelab enthusiasts automate tool research and replacement. Built on the Hermes Agent framework, toolsmith operates with a strict safety-first philosophy, refusing to execute any command it cannot back with measured facts. The system is designed to audit one specific area of a user's homelab each week, identifying potential replacements for existing tools and testing the best candidates in throwaway containers.
The Safety Guard Architecture
The core innovation lies in toolsmith's pre_tool_call hook, which acts as a mandatory gatekeeper for every action the agent attempts. This guard, implemented in Python, intercepts JSON payloads before execution. For example, a command like 'rm -rf /' is immediately blocked with a decision of 'block', while a web search for 'caddy vs traefik' is allowed to proceed. Crucially, the system is wired fail_closed, meaning if the guard script itself crashes, the associated tool call is automatically blocked rather than allowed, preventing silent failures from compromising the system.
Installation and Bug Fixes
The setup process, hosted on GitHub by user casareanderson, recently faced a critical bug that highlighted the importance of rigorous testing. The original README instructed users to run tests before installing, but this sequence caused the installer to fail silently. The tests generated a __pycache__ directory that the installer's copy command couldn't handle, leading to a partial installation without profiles or configs. The fix involved modifying the installer to copy files only, ensuring a clean setup on fresh Ubuntu machines, which are now tested weekly.
Infrastructure Requirements and Limitations
While the safety guard and test suite are portable Python scripts, the full toolsmith experience requires specific infrastructure. It depends on Hermes Agent and Proxmox to perform its primary function of sandboxing tests in real containers. The agent communicates with the user via Discord, requesting approval for changes with simple โ / โ reactions. However, CI pipelines cannot fully validate the end-to-end workflow because they lack the necessary Discord and Proxmox environment variables, stopping validation after the guard tests.
Key Takeaways
- Toolsmith enforces a fail-closed security model where any guard crash blocks execution.
- The agent tests tool replacements in isolated Proxmox containers before asking for user approval.
- A recent installer bug caused by __pycache__ directories was fixed by restricting file copy operations.
- Full functionality requires Hermes Agent, Proxmox, and Discord, though the safety guard runs on standard Python.
The Bottom Line
Toolsmith proves that agentic automation in homelabs can be both ambitious and paranoid, offering a blueprint for safe, fact-based tool management. It is a necessary evolution for anyone tired of AI agents hallucinating destructive commands, prioritizing verified facts over blind execution.