Design partner program is openThree partners, real decision traffic →
razoo

The governance kernel for AI agents

Give your AI only the authority it has earned. Then watch the model bill fall.

Access control decides what your software may touch. Razoo decides what your AI may decide on its own. Nothing touching money or contracts runs without a named person, and the calls your people make the same way every week earn the right to run without the model, until you take that right back.

Not technical? Read the two-minute version for leaders

invoice-under-limitRule

Routes and codes invoices under the delegated limit from known suppliers. Approval stays with a person.

0 tokens · 412 fires

contract-expiryLearned

Flags agreements expiring inside 30 days and drafts the renewal note.

certified by replay · revocable

unknown-senderModel

Classifies mail from unknown parties. Residual goes to the model under a token contract.

capable tier · 1.2k tokens

billing-routingLearned

Routes billing questions to accounts.

certified by replay · week 8

illustrative decisions · one graduates, one is revoked

Works withOpenAIAnthropicGoogleOpenAI-compatibleLangGraphCrewAIGmailIMAP

01 · The problem

Every company running agents faces an impossible choice. Approve everything, forever, and drown in sign-offs. Or trust the model and hope, which no auditor accepts. Either way, the agent that made the same decision ten thousand times this year is no cheaper, faster, or more trusted than on day one. Agents never graduate.

Convert every repeat

Human judgments become evidence. Evidence certified by replay against your own history becomes a rule that runs without the model.
How authority is earned

Bound every authority

A learned rule fires only inside an envelope that names the fields, actions, and limits it may touch. Money and legal fields always need a person.
Inside the kernel

Revoke in one click

A counterexample, a change in what it depends on, or one click, and the rule loses force everywhere at once. Work returns to a person. Nothing is dropped.
Inside the console

02 · Graduation

In plain terms

There is a third option. Graduation.

AI discovers the rule. Razoo proves the rule. Software inherits the rule. Twelve weeks of decisions, drawn: grey ones the model made, green ones a learned pattern made, brass ones a rule made for nothing. In week eight, one pattern gets something wrong and is revoked.

model learned · revocable rule · zero tokens revokedseeded simulation, not customer data
  1. 1

    Your agent keeps working as it does today

    It still uses whichever AI model you already pay for. Nothing about it changes on day one.

  2. 2

    Your people make the calls they already make

    Approve the invoice. Send the reply. Chase the renewal. Razoo records each decision as evidence.

  3. 3

    Repeats stop needing the model

    When your people keep making the same call the same way, and replay against that history certifies it, Razoo makes the call itself, for nothing. If it ever gets one wrong, you take it back with one click.

03 · Where it sits

Where Razoo sits

Under the agent, before any side effect.

Control planes decide what agents may do. Frameworks and models make agents smarter. Your systems of record are where the consequences land. Razoo is the kernel between the agent and the consequence.

04 · The console

Watch authority being earned.

You cannot revoke what you cannot see. The console shows every decision, the rung that resolved it, the evidence behind any learned rule, and the confirm behind each side effect. When a pattern is ready to earn the right to run itself, it says so.

Explore the console

Stage everything, execute nothing

Reasoning stages an action. A named person confirms it. The outbox executes it. That order never changes.

How the kernel decides

Explain every decision

Rung, pattern, the version it fired with, envelope, tokens, cost, and the confirm that allowed it. Replayable bit for bit.

The decision inspector

Learn from the queue, not from a training run

Corrections are captured in flow. No prompts, no labelling, no fine-tuning. The evidence is the work you already did.

How authority is earned

Run under any framework, on any model

LangGraph, CrewAI, a homegrown loop, or a short TypeScript loop. Policy names a tier, fast or capable, never a vendor.

Embed with the SDK
console · reference-contractsillustrative
Queuethis week
  • contract-expiryNorthgate

    proposed · replay in progress

    Learned
  • signature-missingHarbour Build

    fired · 0 tokens

    Rule
  • value-reviewMeridian

    money_legal · a person confirms

    Human
  • incoming-contractUnknown party

    capable tier · 980 tokens

    Model
  • renewal-noteNorthgate

    certified · ready to promote

    Learned
Decisiond_9f3a…c21e
Resolved by
learned pattern
Envelope
action · draft renewal note · confirm required
Authority
valid · re-checked before send
Confirmed by
approver, before send
Tokens
0
Was this wrong?Revoke and correct

Decisions

1,284

Still need the model

31%

Tokens avoided

2.1M

05 · The bill

What it does to the bill

Usage keeps growing. The model bill stops growing with it.

Every decision that has earned its place stops needing the model, so the bill flattens while the work keeps growing. The saving is real. It is the benefit, not the reason. Pick your own spend and see the shape; your actual number comes from your conversion report.

Your model spend today$25k a month
$2k$2.0M+

Month 12, without

$47.5k

Month 12, with Razoo

$13.3k

Saving by month 12

$34.2k/mo

72% of that month's bill

Year one, cumulative

$222k

ramps from zero as rules are earned

Without RazooWith Razoomonthly model spend · illustrative
$0$12.8k$25.6k$38.4k$51.3kM1M2M3M4M5M6M7M8M9M10M11M12$47.5k without$13.3k with Razoo
Illustrative model, not customer data. Nothing is saved in month one; the saving ramps as rules are earned, which is why the year total is less than twelve times the month-twelve gap. Starts from $25k a month of model spend, growing 6% month on month with usage, and a conversion curve that leaves 28% of decisions on the model by month twelve. Your own numbers come from your conversion report.
View as a table
MonthStill on the modelWithout RazooWith RazooSaved
1100%$25k$25k$0
286%$26.5k$22.9k$3.6k
374%$28.1k$20.9k$7.2k
464%$29.8k$19k$10.8k
555%$31.6k$17.2k$14.3k
647%$33.5k$15.7k$17.7k
741%$35.5k$14.4k$21k
836%$37.6k$13.4k$24.1k
932%$39.8k$12.8k$27k
1030%$42.2k$12.5k$29.7k
1128%$44.8k$12.7k$32.1k
1228%$47.5k$13.3k$34.2k
Year$422k$200k$222k

06 · Two ways in

Two ways people use it

Agents in your business, or agents in your product.

Running agents in your business

You have agents. You want the repeats to stop costing you.

Deploy the appliance on your infrastructure, connect one workflow, and let your people make the calls they already make. Invoices, contracts, inbox, support, renewals, and a dozen more.

Building agents into your product

You ship a product with an agent inside. Every customer costs you model spend.

Embed the SDK underneath your product. Cost per customer stops growing with their usage, users get the same fast answer every time, and every action stays safe inside your own interface.

Solo operator or small business? Razoo hosts it for you, nothing to install, free to try.

07 · Observe first

Free, forever

How many of your agent's decisions are repeats? Find out without changing a line.

Point your agent's base URL at Razoo Observe. It relays every request byte for byte, records the decision, and produces a report: which calls repeat, which could run without the model, and what they cost you today.

See how it worksRead-only. Byte-faithful. Hosted, free.
one environment variable · illustrative
# before
OPENAI_BASE_URL=https://api.openai.com/v1

# after: same agent, same model, every decision recorded
OPENAI_BASE_URL=https://observe.razoo.dev/a/<your-account>/v1

# then, whenever you like, in the portal
observe/report · 2,140 decisions · 38% structurally repeatable · 22% certifiable

Every provider

OpenAI and Anthropic surfaces, streaming included

Nothing changes

Requests and responses relayed untouched

Nothing shared

Traces in your own account, yours to export or delete

08 · Appliance or SDK

Appliance or SDK

On your infrastructure, or inside your product.

One container with a persistent volume and your identity provider, or one npm package with your agent folder and your SQLite file. Submit, propose, confirm, drain, explain, and revoke from the interface your users already have. No migration, no vendor lock, no data leaving the tenant.

your product, your UI, Razoo underneath
import {
  bootEngine, submitWork, proposeAction,
  confirmAction, drainOnce, explainOf, clockFromIso,
} from "@engine/sdk";

bootEngine({ root: process.cwd() });
const ctx = clockFromIso(now);

const work = await submitWork({
  work_type: "ap_invoice", source: "erp", subject, ctx,
});
const staged = await proposeAction({
  workItemId: work.id, actionId: "approve", ctx,
});

// still pending. Confirm is a separate act.
await confirmAction(staged.id, ctx);
await drainOnce(ctx);

explainOf(work.id); // rung, pattern, envelope, tokens, who

09 · Security

Security

Enforced by construction, and written for the questionnaire.

The guarantees below are pinned by tests that fail if the guard is removed. Everything on the roadmap is labelled as roadmap.

Authority

  • Nothing executes unconfirmed
  • Money and legal fields always need a person
  • One promotion path, certified by replay
  • Stale authority cannot fire

Data

  • One deployment per tenant, on your volume
  • Secrets encrypted at rest, AES-256-GCM
  • No model call unless policy allows it
  • No telemetry. No content ever leaves your tenant

Audit

  • Every decision explains itself
  • Bit-for-bit replay against the current kernel
  • A person's confirm on every side effect
  • Signed export per work item or period

Transport

  • HTTPS only for connectors
  • DNS-resolved private-address guard
  • Bounded bodies, same-host redirects
  • Secrets by reference, scrubbed from errors

10 · Evidence

Evidence, not adjectives

We publish the mechanism and the benchmark.

No logos we have not earned, no numbers we have not measured. Here is what you can read and run today.

Documentation

Install, quickstart, agent authoring, the SDK and CLI references, appliance deployment, and the security model.
Open the docs

GovBench

A seeded, deterministic harness that scores any stack on conversion, unsafe actions, and revocation after drift. Run it on yours.
Run the benchmark

Technical review

Architecture, invariants, threat model, and red-team history. Shared with design partners under NDA.
What the review covers

A coin worth nothing.

Every decision an agent has earned the right to make costs nothing to make again. Everything else still costs a person's confirm or a model call, and the record says which.

Three design partners. Real decision traffic.

If you run agents that make the same calls every week, we want to govern them and show you what stops needing the model.