ControlSpec — Before an agent acts

ControlSpec — Before an agent acts

Replays recorded tests of exact-action approval for AI agents, run through an NVIDIA OpenShell sandbox, so you can inspect approvals, replays and failures. For teams that review agent safety.

Made by claytonsamples on Hugging Face. These notes are Mindket's, written from its Space page and README; the app's own files and license are on Hugging Face.

About

What it is: This Space is a static replay of recorded evidence for ControlSpec, a layer that checks an exact action before an AI agent takes it. You explore 14 recorded probes that ran through an NVIDIA OpenShell v0.1.2 sandbox, ControlSpec middleware over authenticated TLS, and a synthetic HTTPS purchase target. Nothing here runs live.

Who it suits: It suits people who review how AI agents are controlled, and anyone who wants raw evidence rather than a claim.

Industry | Who uses it | How they use it | Example
Security | Security engineers | Inspect how exact approvals and denials played out | Check what happened when a purchase amount changed
AI platforms | Agent platform teams | Study middleware that sits between an agent and its targets | Read the recovery proof after a middleware outage
Governance | Auditors and reviewers | Download the original observations and logs as evidence | Compare the effective policy with what was recorded
Research | Agent safety researchers | Review replay and missing-proof cases | Look at a direct-IP attempt and how it was handled

You can skip it if you want to run a live agent or need a security certification, since this is a replay of scripted tests.

Where to use it:

• A web browser: open the Space page on Hugging Face and explore the probes.
• Windows, macOS or Linux: download the files and open runtime.html in a browser.

Use case | Who | What they get back
Explore the 14 probes | Reviewers | Recorded results for each approval or denial case
Get the raw records | Auditors | The first observations, the policy in force, the supervisor logs and the proof of recovery
Compare earlier demos | Researchers | Links to the earlier action demo and reference examples

How to set it up: The hosted page needs nothing from you, and no signup or keys are required. Downloading a copy needs Python and pip, plus a browser.

• Python, which brings pip, on Windows, macOS or Linux.
• A web browser to open the page.

1. Install the huggingface_hub package with pip.
2. Download the Space's files into a folder with the hf command.
3. Move into that folder.
4. Open runtime.html in a web browser.

pip install -U huggingface_hub
hf download claytonsamples/controlspec --repo-type space --local-dir controlspec
cd controlspec

How to use it: Open the page and pick a recorded probe. Read what was requested, what the middleware decided and what the sandbox did. The probes cover exact approvals, changed amounts, replay, missing proof, denial of a native path, attempts by direct IP and on the control plane, and a middleware outage. Each probe has downloads of the raw record.

Tool | What it does
runtime.html | The main page with the 14 recorded probes
action-demo.html | A page with an earlier action demo
reference-demo.html | A page with the original reference examples
JSON files | The recorded observations and traces the pages replay

Example: You open the page and choose the probe where the purchase amount was changed after approval. You read how the middleware treated the changed request and what the sandbox logged. Then you download the original observation file to check it yourself.

For newcomers and experts: If you are new, think of it as a flight recorder for an AI agent's actions. Open the page, pick one probe and read its result from top to bottom. Nothing real is spent and no live model is involved.

If you know the field, the runs used scripted requests and test-operator approvals on Docker Desktop with WSL2. The sandbox is NVIDIA OpenShell, and the target is a synthetic purchase endpoint. Selected probes passed. A separate protocol lab keeps the earlier nine scenes, and its native comparison lane has not been run.

A scripted request is made by the agent
ControlSpec middleware checks the exact action
The sandbox allows or denies it
The observation and logs are recorded
You replay the record in the page

Good to know: This is a static evidence replay, with no live model, money or outside action. Do not treat it as a security certification for production or as a benchmark of any platform. ControlSpec is independent, and no endorsement by NVIDIA or Hugging Face is implied.

• License: apache-2
• Cost to run: none named, and no signup or keys are required
• Data it reaches: none named, since it replays recorded files
• Not for: production security certification, platform benchmarking or running live agents

On Hugging Face, 13 November 2026

• Likes: 1
• Built with: static
• License: apache-2.0

Links

• Space: https://huggingface.co/spaces/claytonsamples/controlspec
• Files: https://huggingface.co/spaces/claytonsamples/controlspec/tree/main

Try it

Open the Space in your browser, or copy it to your own Hugging Face account from its page. Nothing in this kit runs by itself.