Skip to content
Get started

Nerve Governance

Authority that doesn't depend on the model behaving

Nerve Governance decides what Agents may do, who may do it, and when a person must say yes. It is enforced outside the model, in the Tool Gateway, so a clever prompt cannot talk its way past it.

Build an AgentTalk to sales

Nerve Governance · one action, four gates

An Agent asks to: Send a quote by email · Risk: Sends externally

  1. 1

    Is this capability allowed?

  2. 2

    Is this identity allowed?

  3. 3

    Is this connection granted?

  4. 4

    Does policy need a person?

Demonstration with sample data

01Capabilities, classified
Every tool Agents can reach has a readable name, a risk level and an approval rule. Administrators can raise risk freely. Lowering it needs a special permission and a written reason.
02Approval policies
Decide which actions run on their own, which need approval and which always do. The rule that decided is shown on the capability.
03Execution identities
Agents act as a defined identity with its own permissions, not as a shared super-user. Managers see and change identities in one place.
04External tool servers, reviewed
MCP servers are reviewed before their tools are available. A changed tool is a new review, not a silent update.
05Nerve MCP access
Control which outside clients may use Nerve over MCP, with the same gates as everything else.
06Activity you can audit
Requests, approvals and executions are recorded with who, what and when. Sections you cannot read say so in a sentence instead of disappearing.

How it is different

Rules the model can't argue with

Telling a model to ask before it acts is a hope. Nerve puts the rule where the model has no vote.

  • Where rules live

    A typical agent platform:In the prompt, as instructions the model is asked to follow.

    Nerve:In the Tool Gateway and the Approval Engine. The server re-checks every call whatever the model says or the screen shows.

  • Default

    A typical agent platform:Tools are available unless someone removes them.

    Nerve:Deny by default. An empty policy admits nothing and there is no allow-all switch.

  • What an approval means

    A typical agent platform:A general yes to a kind of action.

    Nerve:Exactly one execution of exactly the action shown: this tool, these arguments, this requester. Change any of them and it asks again.

  • Lowering a risk rating

    A typical agent platform:Anyone with settings access flips it.

    Nerve:Needs its own permission, never goes below Nerve's floor, and requires a written reason.

  • Who approves

    A typical agent platform:Whoever is online, including the requester.

    Nerve:A policy can require a different person to approve, so the one who asked cannot approve their own request.

  • Evidence

    A typical agent platform:Logs scattered across systems.

    Nerve:One activity trail of requests, decisions and outcomes, with the rule that applied.

A comparison of design approaches, not a benchmark.

Four questions before every action

  1. 1

    Is it allowed?

    The capability must be granted to this Agent, and the identity must hold it.

  2. 2

    Is it connected?

    The connection must be granted for this workspace and this action.

  3. 3

    Does a person decide?

    Policy runs it, asks for approval or always asks. The result is recorded.

Powered by Nerve Cortex. Every Nerve Agent, every Command turn and every workflow step runs on one durable runtime, built by Nerve, that keeps its place, asks before it acts and never repeats an effect.

Meet Nerve Cortex

Deploy your first Nerve Agent

Start with one conversation your team handles every day. Give the Agent what it needs to know, connect the tools it should use, and decide where a person steps in.

Build an Agent Talk to sales