The ollama 0.14-rc2 release is available today and it introduces new functionality with ollama run --experimental for in this experimental mode to run an agent loop so that LLMs can use tools like bash and web searching on your system. It’s opt-in for letting ollama/LLMs make use of bash on your local system and there are at least some safeguards in place.

Sure to rub some people the wrong way, the new experimental agent loop with ollama 0.14 allows for LLMs to leverage some tools like Bash. Bash command execution can be driven via LLMs but for making sure the LLMs don’t go too wild, there is an interactive approval user interface. There is also an auto-allowlist for “safe” commands as well as a deny list for blocking potentially dangerous commands like “sudo” and “rm -rf”. Though among the safe commands are npm run, pwd, git status, and others. The experimental agent loop will also provide a warning box for commands working on paths outside the project directory