Get Sheaf

Three steps. No change to your agent.

A key in a minute. A plan when you need more than the free allowance. Then Sheaf in front of your agent's tools on NVIDIA OpenShell, so every gated action is proved before it happens, held when it does not add up, and recorded.

Get a free keySkip to install

1

Get a key

Your email, and a key comes back at once. No card. The free allowance is enough to connect one agent and see a hold happen.

Create a free account

The key is shown once. If you lose it, the same page issues a new one by email.

2

Know the price

You pay 35 cents for each action your agents send for a full check. Rules checks are free. Your first checks each month are free with no card, and you add a card when you want more. Figures are pulled live from the same place the backend charges from.

Free to start

$0

About 12 full checks a month

One agent, one hold, no card.

Get a key

Running a lot? A committed year costs less per check, and Sheaf can run inside your own systems: pricing.

3

Put Sheaf on the line

Two ways. The hosted gate takes five minutes and needs nothing installed. The self-hosted proxy keeps every tool call on your own machines.

Hosted gate

The setup wizard reads your agent's tools, asks which actions must never fire without a check, drafts the panel that checks them, and hands you one address. That address is a transparent MCP proxy Sheaf runs for you. Put your key on the wizard's first screen and the gate runs on your plan.

Open the setup wizard

Self-hosted proxy

The same proxy, from npm, on your own machine. It holds your real tool credentials and your key; the agent sees neither.

npm i -g @sheaf-labs/mcp-proxy
SHEAF_API_KEY=sk_pod_… sheaf-proxy sheaf-proxy.config.json

The package and its README

On NVIDIA OpenShell: make Sheaf the only door

The five-minute proof, one command, on your own OpenShell:

curl -sSf https://sheaf.one/openshell/quickstart.sh | bash

It creates a sandbox where Sheaf is the only tool endpoint, shows the fence, creates a gate over an example tool server, and from inside the sandbox lists the tools through Sheaf, passes a harmless call, and holds a suspicious payment for a person. Read the script first if you prefer. Run end to end on OpenShell, 28 September 2026.

OpenShell fences what your agent may reach. This policy removes the agent's real tool servers and leaves one endpoint: the Sheaf gate. OpenShell enforces that at the kernel; Sheaf decides what goes through it. Two files, both plain YAML.

# the sandbox policy: the agent's only tool endpoint is the Sheaf gate
curl -sO https://sheaf.one/openshell/sheaf-policy.yaml
openshell sandbox create --name my-agent --policy sheaf-policy.yaml

# prove the fence in one minute (OpenShell's default image has no curl; use one that does)
openshell sandbox create --name sheaf-test --from curlimages/curl:latest --policy sheaf-policy.yaml --no-auto-providers --detach -- sleep infinity
openshell sandbox exec -n sheaf-test -- curl -s https://api.sheaf.one/api/v1/plans       # JSON: Sheaf is reachable
openshell sandbox exec -n sheaf-test -- curl -s https://api.github.com/zen               # nothing: everything else is denied

# the provider profile: OpenShell holds your Sheaf key outside the sandbox and attaches it only to api.sheaf.one
openshell provider profile import -f sheaf.yaml --global
openshell provider create sheaf --type sheaf
openshell sandbox create --name my-agent --provider sheaf --policy sheaf-policy.yaml

↓ sheaf-policy.yaml↓ sheaf.yaml

Tested on OpenShell 28 September 2026. Inside the sandbox, point the agent's MCP config at your gate address from step 3 and it is done. Set binaries in both files to the paths in your image; the policy lists curl and node. OpenShell is an NVIDIA product; Sheaf is independent.

Coming: Sheaf as an OpenShell middleware

OpenShell has a preview extension point that hands a middleware the body of every request leaving a sandbox before the credentials are attached, and lets it deny with a reason. That is the tool call itself. Sheaf as that middleware means the check runs inside OpenShell's own contract, with no proxy to point at. It is being built now. Write to us to be first on it.

Then

Run your agent. The first gated call it makes shows up on your account page, checked, held or released, with the record beside it. When the regulator asks why, you hand over the record.

How the two layers fitHow the signals work