auto-150m-3

auto-150m-3

Decides whether a coding agent's proposed tool call should run, answering approve or deny from the call, the request and the history. The smallest Auto model, for machines with little memory.

Made by ProCreations on Hugging Face. These notes are Mindket's, written from its model page and model card; the model's own files and license are on Hugging Face.

About

What it is: auto-150m-3 is a 149.6M-parameter classifier that checks an AI agent's proposed tool call before it runs. It looks at the call, the request behind it and what the agent did earlier, then says approve or deny. It is the smallest and fastest model in the Auto family, built for machines that cannot spare much memory.

Who it suits: People who run coding agents and want a local guard that blocks risky calls the user never asked for, or calls that follow instructions hidden in a page, file or tool output.

Industry | Who uses it | How they use it | Example
Software | Developers using coding agents | Screen each tool call before it executes | Deny a delete command the user did not request
Security | Agent safety teams | Add a local check against planted instructions | Block a call that follows text hidden in a web page
AI tooling | Agent builders | Put a small approve or deny classifier in front of tools | Run it on a laptop CPU at low cost

Anyone who wants the highest accuracy can skip it, since the larger Auto models score higher.

Where to use it:

• Auto, a local guard for coding agents such as Claude Code, Codex, Gemini CLI, Cursor and Copilot CLI.
• The transformers library, since the record lists it.
• Almost any machine, including older CPUs, because the model is small.
• Windows, macOS or Linux: the download steps run on all three with Python.

Use case | Who | What they get back
Tool call screening | Agent users | An approve or deny decision for each proposed call
Prompt injection defense | Security teams | A deny for calls driven by planted instructions
Low-memory guarding | Developers on small machines | A decision in about 5 ms on the card's test GPU

How to set it up: What you need first:

• Python with pip.
• The transformers library.
• Optionally, the Auto project, which wires the model into coding agents.

1. Install the Hugging Face hub package with pip.
2. Download the model with the hf command.
3. Load it with transformers, or let Auto use it.

pip install -U huggingface_hub
hf download ProCreations/auto-150m-3

How to use it: You pass in the proposed call, the user's ask and the earlier steps. It returns approve or deny, and a call is denied when its deny probability reaches 0.5. The card's usage code is not in the text available here, so check the model page for the exact input format.

Example: A coding agent proposes a command that wipes a folder, and the user never asked for that. The model reads the call, the request and the history, and returns deny. A call the user did request returns approve.

For newcomers and experts: Think of it as a gatekeeper for an AI agent. Before the agent runs a tool, the gatekeeper says yes or no. A good first step is to read how the Auto project connects it to your agent.

The model is a ModernBERT-base encoder with BF16 weights of 0.30 GB. Its test set has 2,948 held-out tool calls. Each has a close twin in which one small detail reverses the correct decision. The trade-off is that it misses 9.1 percent of harmful calls and blocks 13.1 percent of authorized ones, so larger Auto models fit where accuracy matters more.

Agent proposes a tool call
Model reads the call, request and history
Model scores the chance of deny
Call runs on approve and stops on deny

Good to know: The card compares accuracy on its v3 benchmark.

Title: v3 benchmark accuracy
auto-3b-3: 96.0 percent
DeepSeek V4.1 Flash: 95.6 percent
auto-0.4b-3: 93.1 percent
auto-150m-3: 88.8 percent
GPT-6 Luna: 87.9 percent

• Speed and cost: a median of 5 ms per decision on an NVIDIA RTX PRO 6000, and about $13 per million decisions on a rented GPU.
• Older benchmark: 96.0 percent on the original approve-or-deny test.
• Privacy: the card says Auto works locally, so its inputs stay on your computer.
• License: apache-2.0
• Cost to run: no API key, account or paid service named.
• Data it reaches: the tool calls, requests and histories you pass in, and files you download from Hugging Face.
• Not for: cases where missed harm or over-blocking cannot be tolerated at these rates.

On Hugging Face, 13 November 2026

• Task: text classification
• Library: transformers
• Parameters: 150M
• Weights: about 299 MB in BF16; about 75 MB at 4-bit (estimate)
• Formats: Safetensors
• Downloads in the last 30 days: 207
• Likes: 4
• License: apache-2.0
• Based on: ProCreations/auto-200m-2
• Quantized versions: 1
• Tags: transformers, safetensors, modernbert, text-classification, agent-safety, tool-calling, guardrail, long-context, model-index, text-embeddings-inference

Links

• Model: https://huggingface.co/ProCreations/auto-150m-3
• Files: https://huggingface.co/ProCreations/auto-150m-3/tree/main

Get the model

Download its files with Hugging Face's command line:

pip install -U huggingface_hub
hf download ProCreations/auto-150m-3

Then follow the usage in its model card. Nothing in this kit runs by itself.