Arbiter

Stop unauthorized agent actions — without rewriting or redeploying your agents.

The Control Plane for AI Agents

Every action reaches Arbiter before production — Allow, Hold, or Deny. Humans approve when required, and every decision becomes audit evidence.

Fastest start: MCP Authority Gate — zero-code governance for MCP-connected agents.

Become a Design PartnerOpen Workspace
arbiter — product demo

Works with the AI agent frameworks you already use

Keep your frameworks. Connect through MCP, SDK, CLI, or API.

The risk

Your agents can already move money and break production.

Payments, infra changes, customer data — without a shared stop before execution, you find out after the damage.

Happening now

Finance Agent

Sends a $50,000 payment

No human in the loop

Happening now

Engineering Agent

Deletes a production database

No review. No record.

You need a stop before execution — not another policy buried inside each agent.

Proof

Watch one action get stopped — then approved.

A $5,000 payment request. Arbiter holds it. A human decides. The receipt is permanent.

AgentActionEvaluateHoldApprovalAudit

Governance Flow

Paused

Agent Identity

finance-agent

Autonomous payment processing agent

Exploring manually — auto-play resumes after a pause

How Arbiter works

That loop is the product.

Not a pile of tools. One path every action takes before it can touch production.

  1. 1

    Agent

    Arbiter knows which agent is acting — not just which model answered.

  2. 2

    Action

    The agent asks to do something real: pay, delete, export, deploy.

  3. 3

    Runtime Evaluation

    Before anything executes: Allow, Hold, or Deny.

  4. 4

    Human Approval

    Hold pauses the action until a person authorizes it.

  5. 5

    Execution Release

    Only then can the action reach Stripe, GitHub, your DB, or your APIs.

  6. 6

    Audit Evidence

    You can explain every decision later — with a receipt, not a guess.

arbiter — policy registry
Arbiter Policy Registry with allow, hold, and deny decisions, conditions, and priorities
Where Allow / Hold / Deny is defined. Rules run at evaluation — before the action executes.
arbiter — governance audit trail
Arbiter Governance Audit Trail with evaluation records, decisions, and receipt links
Where every decision lands. Trace who asked, what was decided, and whether a human released it.

Connect your stack

Keep OpenAI Agents, LangGraph, CrewAI, MCP — add a stop before production.

Founder-led path: start with the MCP Authority Gate. Use the SDK when you want evaluation inside your agent code.

MCP Authority Gate

Recommended start. Point your MCP client at Arbiter — govern tools without rewriting agents.

SDK

Call evaluate() from agent code. Get Allow, Hold, or Deny before you execute.

Manifest & CLI

Declare agents in arbiter.yaml and sync posture from the terminal.

REST API

Evaluate, approve, and fetch audit evidence over HTTP when you need full control.

SDK path — one evaluate call

import { Arbiter } from "@arbiterhq/sdk";

const decision = await arbiter.evaluate({
  agent: "finance-agent",
  action: "send_payment",
  context: { amount: 5000 },
});
// decision: Allow | Hold | Deny

What you get

Operational control — in plain terms.

Five things between your agents and production. Nothing more to learn on the homepage.

  • 1

    Who is acting

    Every agent has an identity you can authorize and revoke.

  • 2

    What it may request

    You define which actions each agent is allowed to ask for.

  • 3

    What happens at runtime

    Allow, Hold, or Deny — decided before anything executes.

  • 4

    When a human decides

    Sensitive actions wait for a person — with full context.

  • 5

    What you can prove later

    A receipt for every decision. Not a chat log guess.

Built with Developer Feedback

We asked developers, founders, security engineers, and enterprise reviewers to challenge Arbiter. Here's what they said—and what we're fixing.

What early developers are saying

  • “From sign up to a real $5,000 approval flow in under 30 minutes—with zero prior knowledge—was genuinely impressive. The AI policy generator was the cherry on top.”

    Khushi

    Early Reviewer · LinkedIn

  • “The SDK integration was straightforward. A centralized place for agent governance—without changing application code—is a strong concept.”

    Harshit

    Developer · LinkedIn

  • “Decoupling identity, authorization, policy evaluation, approvals, and auditing into an external control plane makes the architecture more modular and secure.”

    Misbah Ullah

    Co-founder, GADIS · LinkedIn

  • “A very thorough governance system with clear separation between different parts of the workflow.”

    Ryan Gormican

    Product Reviewer · LinkedIn

  • “'Without changing code or redeploying them' is the reason a busy team would even try you.”

    Dima

    Founder, NoCrash · LinkedIn

  • “A dedicated security gateway adds an independent layer of control before agent actions reach production systems.”

    Deekshitha

    Security Reviewer · LinkedIn

  • “The vision is compelling. The positioning and architecture are clear.”

    Phani

    Founder, Zyvora · LinkedIn

  • “The Allow / Hold / Deny model and the incident examples immediately make the problem obvious.”

    Jogeeshwara

    Advisor · LinkedIn

  • “Arbiter shows a strong, coherent approach to AI agent governance—with clear technical direction and enterprise-oriented thinking.”

    Bruno Baumgartner

    Founder, DigiEmu · Direct Outreach

We're listening — here's what we're fixing

  • We heard

    “Too much to grasp at first. I didn't find enough code snippets to understand how it works.”

    Divakar

    AI Developer · LinkedIn

    Improving

    • Framework quickstarts (LangGraph, CrewAI, MCP)
    • Simpler first-run experience

  • We heard

    “Onboarding was confusing. Errors like identity_pending_review didn't tell me what to do next.”

    Harshit

    Developer · LinkedIn

    Improving

    • Guided onboarding
    • Action-oriented setup errors

  • We heard

    “The landing page undersells the product. Terminology was confusing and sections felt disconnected.”

    Ryan Gormican

    Product Reviewer · LinkedIn

    Fixed

    • Guided homepage story
    • Problem → proof → runtime → adoption

  • We heard

    “You lead with the category, not the fear. Policy Registry and Permission Registry sound like the same thing.”

    Dima

    Founder, NoCrash · LinkedIn

    Fixed

    • Lead with unauthorized agent actions
    • Permissions vs policies in the runtime story

  • We heard

    “The dashboard immediately shows setup errors. There's no guided demo mode.”

    Abdul

    Technical Reviewer · LinkedIn

    Improving

    • Clearer first-run workspace experience
    • Reducing setup-error noise on entry

We're an early-stage company. Trust comes from listening to developers, shipping improvements, and being transparent about what's still being built—not from pretending the product is perfect.

Build this with us.

We work directly with a small set of teams shipping agents to production. Early access, founder and engineering support, and real influence on what ships next.

A fit if you are

  • AI Startups
  • Agent Builders
  • Enterprise AI Teams
  • AI Consultancies
  • Workflow Automation Platforms

Apply for Design Partnership

Tell us how your agents reach production. We'll respond personally.

  • Early access to new capabilities
  • Direct engineering support
  • Influence the roadmap
  • Priority onboarding
  • Control-plane architecture guidance
Apply for Design PartnershipProgram details