The_Swarm
A local-first LLM tool-calling harness with workspace isolation and safety gates.
The brief
Most agent harnesses either lock you into a cloud provider or run tools with too little isolation. I wanted a local-first harness where the model can plan, write, test, analyze, and remember — but only inside a safe, isolated workspace.
What I made
I built The_Swarm as a conversational CLI with extensible tools (plan / write / test / analyze / memory), a provider abstraction that works with Ollama and OpenAI-compatible endpoints, workspace isolation, and safety gates. It ships as a one-command install for Linux, macOS, and WSL.
Reflection
- Local-first is a real differentiator — most harnesses assume a cloud provider.
- Safety gates and isolation are what make an agent harness trustworthy enough to run tools at all.
Tech stack
- Python
- Ollama
- CLI
- agents
- tool-calling
Results
- A working local-first agent harness you can install and run in one command.
- Provider abstraction means no cloud lock-in — Ollama or any OpenAI-compatible endpoint.
- Workspace isolation and safety gates make tool execution safer by default.
I need something built.