Shepherd
FeaturesBlogAboutDownload for

The governance layer for organizational AI

Shepherd tells your AI agents what they are authorized to do before they act, brings a person in when a limit is reached, and signs every answer into a record you can prove.

Download for

Inbox

Waiting on you
Approval requiredsupport-agent-2 · 6 minutes ago

Refund 150.00 to customer c_4821

  • Authorized up to 150.00 for a silver-tier customer.
  • Above 100.00, a support lead approves.
ApproveDecline
Recent answers
AuthorizedSend email12s
Authorized with limitsRefund 64.001m
Not authorizedForce-push to main4m
Approval requiredRefund 150.006m

Works with the Claude Agent SDK, the OpenAI Agents SDK, LangGraph, any MCP tool and any HTTP client.

Authorize

Authorize any action before it happens

Emails, files, commands, payments and the other tools your agents call. Before an agent acts, Shepherd checks the action against your rules and what has already happened, then answers.

  • Email
  • Files
  • Commands
  • Payments

outreach-agent asks to

Email a discount code to 1,200 customers

Checking against your rules…

Approval required

Emails to more than 100 customers wait for the marketing lead.

ops-agent asks to

Open finance/payroll-2026.csv

Checking against your rules…

Not authorized

Payroll is outside the folders this agent may read.

code-agent asks to

git push origin fix/login

Checking against your rules…

Authorized

Pushes to any branch but main are authorized.

support-agent-2 asks to

Refund 180.00 to customer c_4821

Checking against your rules…

Authorized up to 150.00

Silver-tier refunds stop at 150.00, and the team has 150.00 left today.

Prove

A record that proves itself

Every decision and its outcome is signed and chained to the one before, so any change after the fact shows, and an auditor can check it without taking our word for it.

  1. AuthorizedSend email

    receipt 4809, signedchain intact

  2. AuthorizedRefund 150.00Refund 15.00

    receipt 4810, signedchain intactedited after signing

  3. Not authorizedForce-push to main

    receipt 4811, signedchain intactlink to 4810 broken

  4. Approval requiredRefund 220.00

    receipt 4812, signedchain intact

Chain verified

Chain broken at receipt 4810

Adapt

Write rules the way you would say them

Describe a rule in plain English, read it back, and see what it would have done to last month's real decisions before a named owner makes it live.

You write

Refunds over 100 need a support lead. Nobody gets more than 3 refunds in 30 days.

Shepherd reads it back

  • A refund above 100.00 waits for a support lead.
  • A 4th refund within 30 days is not authorized.

Tested on September

31 of 1,204 decisions change

Live

Accepted by the head of support

Oversee

See every agent, and stop any of them

Every decision your agents ask for, as it happens, and one switch to pause a single agent or the whole fleet.

ActivityPause all

  1. Pay invoice 2,400.00Approval required

    billing-agent

  2. Force-push to mainNot authorized

    code-agent

  3. Refund 180.00Up to 150.00

    support-agent-2

  4. Send 12 emailsAuthorized

    outreach-agent

  5. Refund 60.00Authorized

    support-agent-1

  6. Delete build/Authorized

    code-agent

All agents paused. Approvals already waiting are kept.

Comply

Know where you stand on every framework

Your agents are checked continuously against the frameworks you answer to and the policies you wrote, with the evidence already in the record.

Compliance status, all agentschecked continuously

  • EU AI Act14 of 16 controls passing

  • NIST AI RMF22 of 24 controls passing

  • ISO/IEC 4200131 of 38 controls passing

  • Your own policies6 of 6 controls passing

AI that acts within the authority you gave it.

Bring Shepherd to your next agent.

Download for