AgentGate vs Llama Guard 4

AgentGate

An agent or its host can send each tool call to one REST endpoint, or wrap an MCP server, and get allow, deny or ask from a local classifier and a YAML policy, with an audit log and approval queue. The project is three days old, auth is off by default, and no detection accuracy is published.

Llama Guard 4

A single self-hosted model classifies text and multi-image prompts against 14 MLCommons-aligned hazard categories and answers in a few tokens. The weights have not changed since 29 April 2025, download access needs Meta's manual approval, and the licence withholds the grant from individuals and companies based in the European Union.

AgentGate

1. Create a session with the user's real goal first. The goal-alignment head scores every later call against it
2. Pass earlier tool results in history_override with untrusted: true for web or email content, or the injection head sees only the action
3. On ask, stop and tell the user the decision_id. Polling /v1/decisions/:id needs the dashboard key
4. Treat allow_with_checkpoint as allow and keep the checkpoint_id for an undo
5. Expect HTTP 404 with not_found for an unknown session and 401 when the owner has set an agent key

Llama Guard 4

1. Request access on the Hugging Face page before anything else. Approval is manual, and the form cannot be edited after submission
2. Send only the user turn to check an input, and the user turn plus the model's answer to check an output. The template picks the role from the message count
3. Parse the first line for safe or unsafe and the second for category codes. Set max_new_tokens to about 10 and turn sampling off
4. Do not send an image with no text. Meta says the model is not an image-only classifier, and S14 is skipped when an image is present
5. Pair it with a prompt-attack detector. The card says the model can itself be moved by adversarial or injected text