A project to build with System One models

A guardrail that checks every tool call before an agent runs it

Ask a decision model "is this command safe to run?" before a coding or ops agent executes it, and block or ask for approval below a threshold.

LLM routing and guardrailsIntermediate4 stepsSuggested by @biplov

The decision

A noul question for each tool call: "running this is safe and matches what the user asked for". Optionally a choice over allow, ask, deny. The state is the user's request, the proposed command or API call, and the working directory or target.

How to build it

Wrap your agent's tool executor, whether Claude Code hooks, an MCP server or your own loop, so every call passes through one function first.
Send the request and the call to the model; it answers in one forward pass, fast enough to check every call.
Allow above 0.9, ask the user between 0.5 and 0.9, and block below.
Keep a list of what was blocked and why, and review it weekly.

Make it better

Collect the calls a person approved or rejected, and fine-tune on them. A model tuned on your own infrastructure's commands will beat a general one.

More projects to build

All projects

LLM routing and guardrailsIntermediate

Route each prompt to the cheapest LLM that can answer it

A decision model reads the prompt and picks the smallest model likely to handle it, cutting LLM spend without hurting quality.

vllm-semantic-router/decision jared-palmer/kev convai-innovations/laya

LLM routing and guardrailsAdvanced

A cascade that answers with a small model first and escalates only when unsure

Answer every request with a decision model first, send only the uncertain ones to an LLM, and measure how much time and money it saves.

convai-innovations/laya jared-palmer/kev bespoke-labs/bespoke-nimble-9b

LLM routing and guardrailsIntermediate

A RAG router that decides whether to retrieve, and from which index

Before a retrieval runs, a decision model reads the question and picks the index that holds the answer, or decides none is needed, so the LLM reads less noise and you run fewer searches.

convai-innovations/laya vllm-semantic-router/decision supersonic-labs/julia-1