NXBrain — AI gateway and governance layer for every model

AI gateway · governance layer for every model

Control every AI call.
No surprise bills.
No leaked data.

NXBrain sits between your apps and every AI model, cloud or self-hosted. It caps what each project or client spends, stops sensitive data before it leaves, and keeps a tamper-evident record of every call: any change shows, and any client or auditor can verify it on their own.

Private beta · Free until launch · Your model keys, no markup · Hosted in the EU

  1. 01

    Control

    Budgets, guardrails, agent limits and model routes act on every call, before it reaches a model.

    Costs, data and agents
  2. 02

    Prove

    Every call is sealed and timestamped. A client or auditor can check the record without having to trust us.

    How verification works
  3. 03

    Comply

    AI inventory, EU AI Act risk and ISO/IEC 42001 evidence, filled in from what actually runs.

    See the compliance workspace

Works with the models and tools you already use. Run NXBrain as your gateway, or keep the one you already have (Portkey, LiteLLM, Cloudflare): it keeps executing the calls, and NXBrain sets the rules, meters the cost and seals the record.

Ship it this afternoon

Two lines change. Your code doesn't. Point any OpenAI-compatible SDK at NXBrain, and every call is governed, metered and sealed.

In every response

Works anywhere you can set an OpenAI base URL: official SDKs, LangChain, LlamaIndex, n8n.Read the quickstart

01 · Control · costs

Budgets that actually hold

Set a spending limit per project, team or client. When it runs out, calls stop with a clear message instead of a surprise invoice.

Budgets · October 2026

4 projects · resets Nov 1
support-bot€412.80 of €500.00
83%alerts sent at 50% and 75%
client-portal€1,180.00 of €2,000.00
59%on track
sales-rag€96.15 of €300.00
32%on track
hr-assistant€500.00 of €500.00
Stopped14 calls rejected

The next call to hr-assistant

2:41 PM
HTTP/2 402 Payment Required
{
  "error": {
    "code": "BUDGET_EXCEEDED",
    "project": "hr-assistant",
    "limit_eur": "500.00",
    "message": "This project has used its October budget (€500.00). Request an increase or wait until November 1."
  }
}
  • Limits per project and per agent
  • Alerts at 50%, 75%, 90% and 100%
  • Metered to the millionth of a euro

01 · Control · data

Control on every call
before anything reaches a model

Guardrails, routing and budgets run in the request path, so your rules apply to the call itself, not to a report a week later.

Add:
Redacted, then sentpolicy support-bot/prod

What the model receives

Hi, I'm Dana Whitfield. Please update my account and send the confirmation to [email], or call me at [phone].

  • redacted email
  • redacted phone

Names pass in this policy because support needs them. Name detection can be turned on per project.

Guardrails

Detect personal data, payment details, secrets and prompt injection. Observe, redact or block, per project: your rules decide what leaves.

Explore guardrails
  • Sealedbatch #48213

    1,204 events · root 9f2c1a7e…d81b41e

  • Allowed€0.0042

    sales-rag · nxbrain/fast → openai/gpt-4.1-mini

  • Redactedpii.email

    support-bot · nxbrain/reasoning → anthropic/claude-sonnet-4

  • Blockedprompt injection

    hr-assistant · policy hr/prod

  • Failoverprovider timeout

    nxbrain/reasoning · retried on google/gemini-2.5-pro

Activity log

Every decision, route and cost in one place. Filter it by project, model, gateway or status, and export it to CSV.

See the log

01 · Control · agents

Agents with limits
and an off switch

Each agent gets its own keys, the models it may use, the tools it may call and a budget. Sensitive tools wait for a person, and one click stops the agent everywhere.

invoice-agent

Running
Models
nxbrain/reasoning
Keys
2 active
Per run
40 steps · 10 min · €2.00
Budget
€118.40 of €300.00

Tools

  • read_invoiceAllowedlow risk
  • lookup_vendorAllowedlow risk
  • send_paymentNeeds approvalhigh risk
  • delete_recordsDeniedhigh risk

Waiting for approval

expires in 15 min

invoice-agent wants to run send_payment

{ "vendor": "Northwind Supplies", "amount_eur": 4800.00, "invoice": "INV-20931" }
  1. Step 12 of 40 · read_invoice INV-20931
  2. Step 13 · lookup_vendor Northwind Supplies
  3. Step 14 · send_payment waiting for a person

01 · Control · models

Change the model once
not in every agent

Your apps and agents call an abstract model, like nxbrain/reasoning, instead of a real one. Behind it sits a versioned route: which real models, through which gateways, in what order. To change it, you publish a new version: a second person approves it, critical models must pass your own tests first, and every app picks it up at once. Nothing to rewrite, nothing to redeploy.

Calling nxbrain/reasoning

14 apps and agents
support-botappclaude-sonnet-4
contract-reviewagentclaude-sonnet-4
sales-ragappclaude-sonnet-4
invoice-extractoragentclaude-sonnet-4
research-agentagentclaude-sonnet-4

and 9 more

Route for nxbrain/reasoning

protected · tests required

v7Live

  1. 1anthropic/claude-sonnet-4Portkey
  2. 2openai/gpt-4.1LiteLLM · fallback

v8Draft

  1. 1qwen3:8blocal · Ollama
  2. 2anthropic/claude-sonnet-4Portkey · fallback
  1. Draftmaria@
  2. Validatepending
  3. Testpending
  4. Approvepending
  5. Publishpending
  1. maria@acme.com proposed route v8: qwen3:8b on the local Ollama server first, claude-sonnet-4 as fallback.
  • Fallback to the next model if a provider fails
  • Simulate a route before publishing it
  • Restore any earlier version as a new one

02 · Prove · for security and compliance teams

Proof anyone can verify

If a client, a regulator or a court asks what your AI did, you can show them, and no one has to simply trust you or us. Every event is hashed. Every five minutes the batch is sealed into a Merkle tree and timestamped by an independent authority. Your auditor checks the math, not our word.

  1. Hash every event

    Each call's decision, route, cost and policy become a SHA-256 leaf. No prompt content required.

  2. Seal every five minutes

    Leaves combine pairwise into one Merkle root, following RFC 6962, the scheme behind Certificate Transparency.

  3. Timestamp it independently

    A time-stamping authority signs the root under RFC 3161, so nobody can backdate it. That includes us.

  4. Verify without an account

    Check any report at nxbrain.ai/verify, or offline with the standalone nxbrain-verify CLI.

terminal
$ nxbrain-verify report-2026-10.pdf
✓ 12,418 events in 96 sealed batches
✓ every Merkle root matches its seal
✓ RFC 3161 timestamps valid (last: 2026-10-29 10:20:00 UTC)
Report is intact. Nothing was added, removed or changed.

Live demo · 8 events, one sealed root

sealed root computing…

Press Verify event to recompute the path from this event to the sealed root.

Hashes are computed in your browser with SHA-256.

Private and governed
by design

  • No prompt storage by default

    We record decisions, costs and hashes. Prompts and responses stay out unless you turn on encrypted capture.

  • Your keys, no markup

    Bring your own OpenAI, Anthropic or Google keys. We never resell tokens or take a cut.

  • Isolated by design

    Every organization gets its own encryption key, and data is separated at the row level.

  • NXBrain runs in the EU

    The service, its logs and its evidence are hosted in the European Union. Calls then go to the model providers you choose: use EU regions or self-hosted models to keep them in the EU too.

  • Two-person approval

    Changes to prompts and policies go live only after a second person approves them.

  • Keys you can revoke

    One key per app, team or agent, with expiry and instant revocation.

  • Built for the request path

    Guardrails, budgets and sealing run inline and add milliseconds, not seconds.

  • NXBrain is working toward ISO/IEC 27001 and 42001

    Our own certification is in progress for both standards. It is separate from the ISO/IEC 42001 workspace below, which helps you prepare yours.

  • Leave with everything

    Export your full sealed history in an open format. It stays verifiable without us.

03 · Comply

Compliance without the spreadsheets

Inventory every AI system, classify its risk, and produce the documents auditors ask for, filled in from what actually runs.

AI inventory

5 systems · 2 high risk · 1 action due

SystemPurposeOwnerRisk · indicativeStatus
Support chatbotCustomer supportCXLimited · transparencyDocumented
CV screeningHiringPeopleHigh · EU AI Act Annex IIIIn review · 2 people
Contract summarizerLegal reviewLegalMinimalDocumented
Credit pre-scoringLendingRiskHigh · EU AI Act Annex IIIAssessment due Nov 15
Sales copilotSalesRevenueMinimalDocumented

We don't sell fear of regulation. We help you answer three questions: what data leaves, what it costs, and how you prove it.

The NXBrain principle · by NxHumans

Built for the teams
that build and run AI

Developers & product teams

Ship AI features without surprise bills

Change one base URL. Every call gets a budget, guardrails and a record, and switching models never means a new deploy.

  • One OpenAI-compatible endpoint
  • A key and a budget per project
  • Abstract models, no redeploys
Read the quickstart

Integrators & tech consultancies

Keep each client’s AI separate

Give each client its own project or organization, with its own keys, budget, costs and data, and a record of every call to show them.

  • A project or organization per client
  • Costs and budgets split by client
  • Each organization’s data under its own key
See all features

IT & security

Know what AI your company runs

See which models each team uses, what they cost and what data reaches them. Set limits without blocking anyone’s work.

  • One gateway for every team
  • Hard budgets per project and team
  • Personal data redacted or blocked before it leaves
Request access

More use cases

Compliance & risk

Classify risk and document compliance

Inventory every AI system, classify it against the EU AI Act and keep obligations, impact assessments and evidence up to date from what actually runs.

  • Indicative AI Act risk classification, approved by two people
  • Obligations with owners and due dates
  • Sealed evidence packs for your auditors
See compliance

Delegated administration

Let each department run its own AI

Give every department its own projects, keys, budget and allowed models. Teams manage their day to day, while central IT sets the rules and sees everything.

  • Admin, AI operator, FinOps, auditor and viewer roles
  • Budgets and model policies per department
  • Access requests decided by someone else
See roles and policies

Model quality

Evaluate models as they change

Run your own evaluations on a schedule or from CI. A protected route only moves to a new model when it meets your minimum score, and traffic shifts to the next model when a provider fails.

  • Scheduled evaluations on your own cases
  • Promotion gates: minimum score, maximum regression
  • Automatic failover and conditional routing
See versioned routes

Who's behind it

Built by NxHumans

NXBrain is a product of NxHumans, an AI and process automation consultancy working with organizations across Spain and Latin America. We put AI into production for our clients, and NXBrain brings together the controls those projects need: what each call costs, what data it carries and how to prove it.

  • Spain and Latin AmericaWhere we work
  • Hosted in the EUService, logs and evidence
  • ISO/IEC 27001 and 42001Our own certification, in progress

Simple, published pricing

NXBrain is in private beta. Prices are public, but nothing is charged until launch, and you'll get 30 days' notice before billing starts.

Private beta

We're letting teams in in small groups. Request access, tell us what you want to govern, and we'll reply within 5 business days.

Basic Open in beta

€10€0during the beta

Then €10 per month

Try it and build your AI inventory.

  • 3 AI systems · 2 users
  • 10,000 events a month
  • 7-day retention
  • Gateway with your own keys
  • Basic guardrails
  • Inventory and risk classification
  • Sealed log and public verification
  • PDF reports with watermark
Request access

Starter At launch

€79per month

Not on sale during the beta

Your first AI in production.

  • 10 AI systems · 5 users
  • 250,000 events, then €20 per 100k
  • 30-day retention
  • Budgets, quotas and prompt catalog
  • All guardrails and custom rules
  • Obligations, impact assessment, AI policy
  • Reports without watermark
Request access

Pro At launch

€399per month

Not on sale during the beta

Daily AI use, or a high-risk system.

  • 50 AI systems · 20 users
  • 2M events, then €15 per 100k
  • 12-month retention
  • Agents, evals and conditional routing
  • Opt-in encrypted capture
  • ISO/IEC 42001, technical docs, incidents
  • Evidence packs, RFC 3161, webhooks
  • SSO
Request access

Enterprise

Custom

Regulated industries, built to fit. Annual contract.

  • Custom limits and retention
  • Dedicated deployment
  • Custom detectors
  • SSO, SLA and dedicated support
  • Assisted review
Talk to us

During the beta, approved teams start on Basic. If you need higher limits, tell us in your request.
An event is one AI call that passes through the gateway. An AI system is each app or agent you register in your inventory. Retries and failovers within the same call count as one event.

Start with one URL.

Private beta · Free until launch · Export everything, anytime

Request access

NXBrain is in private beta. Tell us about your team and what you want to govern. We review requests in small groups and reply to every one within 5 business days.

Where are you with AI? · optional
What do you want to govern first? · optional

We only use these details to review your request and reply to you.