AI runtime infrastructure for X-system work

Your AI can think. UniversalBench does the work: one runtime that carries X-system work, the tasks that cross several apps, accounts and sites, from request to finished result. You ask once.

Start free See it work 1,000 free actions a month. No card. What’s X-system work?
Works across your stack
EmailFiles and docsSpreadsheetsAccountingHelp deskCodeDatabasesany websiteAll integrations →
See how UniversalBench compares with Composio, Zapier and others →
WORKINGFINISHED
The request

The result
Supplier chased and delivery confirmed
Invoice adjusted in accounting
Customer emailed the new date
Everything logged for next time
6 systems · 1 orchestrator · 0 handovers Illustrative

Code that runs

A sandbox where the AI calculates, cleans data and builds, instead of only describing.

A browser of its own

Reaches the systems that have no connection option, like supplier portals and older tools.

Your apps, connected once

Sign in with OAuth. No passwords passed around and no integration code to write.

Memory that carries over

Context follows the work from one step, and one session, to the next.

Up to 96.5%fewer tokens on data tasks
1 URLconnects any AI to your stack
1,000free actions a month
Driven by

Behind UniversalBench is Intralace™, our runtime infrastructure. Think of an intranet, an internal system, laced together: it takes the tangle of systems an AI has to work across, handles the complexity inside one controlled place, and hands back one clean result.

How the runtime works

What’s X-system work?

Real work doesn’t live in one app.

A single task usually crosses several systems, each with its own login, its own rules and its own gaps. That’s X-system work, and it’s where most AI tools stop.

One task, six systems

“A customer emails to say their order is late. Chase the supplier, fix the invoice, tell the customer and log it.”

  1. 01Read the emailEmail
  2. 02Look up the customer and their orderAccounting
  3. 03Check the supplier’s site for the deliveryA website with no connection option
  4. 04Correct the invoiceAccounting
  5. 05Reply to the customerEmail
  6. 06Log what happenedA notes app or a spreadsheet

Where it goes wrong

Here’s where a task usually falls apart.

Six systems means six places for the work to stall. Most AI tools stop at the first one and hand the job back to you. This is what the runtime does at each.

01

Every app wants its own login.

So you paste passwords into a chat, or you end up doing that step yourself.

Sign in to each app once. Your passwords stay locked in an encrypted vault, and the AI never sees them.

02

Some systems have no way in.

A supplier portal, an older tool, a website with a login and nothing else.

A built-in browser opens the website, signs in like a person would, and reads what it needs.

03

The steps don’t stay in order.

Real work changes halfway through. The supplier says Friday, not Tuesday.

Each step is checked as it finishes. When the answer changes, the plan changes with it, and the job keeps going.

04

Context gets lost between steps.

By step five you’re explaining the order all over again.

The details travel with the job, from the first step to the last and into next time, so you never repeat yourself.

Inside the runtime

Your AI decides. The runtime does the in‑between.

Signing in, running the code, opening the website, remembering what happened, checking the result. That is the work between a request and a finished job, and it is what UniversalBench carries, so your AI only gets the answer.

you askYour AIClaude, ChatGPT, Gemini, Cursor or any MCP client
the runtimeUniversalBenchDriven by Intralace™. Plans, runs, checks and remembers.
it acts onYour systemsApps, accounts, websites and data you already use
Watch one real task, start to finish“A customer’s order is late. Chase it and sort it out.”
Step 1 of 8Reads your request and plans the steps.
routerRouterReads the request and sends each step to the right part.
connectConnectYour apps, signed in once through their own sign‑in.
workersWorkersIndependent parts of a task run at the same time.
securitySecurityEvery action is checked before it runs, not after.
memoryMemoryContext follows the work from step to step and session to session.
databaseDatabaseReads, writes and searches your data where it lives.
sandboxSandboxCode actually runs, so numbers are calculated, not described.
browserBrowserReaches websites and portals with no other way in.
See how it compares with Composio, Zapier and others Read the docs

Worked out, not guessed

Let your AI think. Let the runtime do the maths.

When a model reads raw data row by row, it guesses, and it burns tokens doing it. UniversalBench runs the calculation inside the runtime and hands your AI only the result. Three tests, run against the live Anthropic API, show the difference.

test 01

Messy CSV revenue

80 sales rows, three date formats, missing and placeholder amounts. Total Q3 2025 revenue from EU customers?

Without2,971input tokens
With UniversalBench101input tokens

Right answer, 25x cheaper output. 96.6% fewer input tokens.

test 02

Statistics over 200 transactions

How many amounts are above $5,000, what is the median, and what is the standard deviation?

Without5,000output tokens, hit the budget
With UniversalBench84output tokens, all three exact

Never finished, versus done instantly.

test 03

Server log error count

A 500-line server log with a realistic mix of INFO, WARN and ERROR. How many ERROR-level lines?

Without20,226input tokens, answered 96 (true: 80)
With UniversalBench48input tokens, answered 80

20% wrong versus correct. 99.8% fewer input tokens.

How we tested. Run on 20 May 2026 with claude-opus-4-7 via the Anthropic API directly, at $15 per million input tokens and $75 per million output tokens. Each “with UniversalBench” call sends Claude a Python-computed answer instead of the raw data, and token counts come from the API’s usage field. The test data is published, and every correct answer is elementary maths you can check in Excel, R or a calculator. Savings depend on the workflow: small queries save less, while bulk data tasks like these save 95% to 99.8%. The “up to 96.5%” figure is a conservative reference from our original public test.

Free tool: LLM cost calculatorSee the tokens and dollars you’d save in a year.Open the calculator →

Guardrails

Limits the AI can’t talk its way around.

Asking an AI to be careful is a request. These are rules the runtime enforces before anything runs, whatever the model decides.

code safety

It never ships broken code.

Every code push is validated before it lands.

  • Syntax validation
  • Live URL smoke test (optional)
  • Automatic rollback support
  • Production-safe deployments
cost guardrails

It never burns your wallet.

Every model call is budget-checked before it runs.

Default ceiling$0.50 / requestHard platform cap$50.00 / requestOver budgetRejected before runSurprise invoicesImpossible
network isolation

It can’t reach your internal network.

Every outbound request is inspected before it goes.

  • Private IP ranges blocked
  • Loopback addresses blocked
  • Link-local ranges blocked
  • Cloud metadata endpoints blocked
asked in a prompt
enforced by the runtime
“Please don’t break anything.”
Broken deployments are blocked before commit.
“Please don’t overspend.”
Overspending is impossible.
“Please stay out of internal systems.”
Internal systems are unreachable by design.
Secrets vaultCredentials are stored encrypted. Never passed to your AI, never logged.
Network isolationThe runtime cannot reach internal IPs or cloud metadata. Enforced at the network layer, not the prompt.
Execution isolationEvery customer runs in its own sandboxed process. No shared state, no cross-customer access.

None of this depends on prompt instructions, agent behaviour or the model choosing to comply.

Getting started

Three steps, then you just ask.

step 01

Create your account

Thirty seconds, no card. 1,000 free actions every month, enough to try it on real work.

your own account and key
step 02

Connect the AI you already use

Claude, ChatGPT, Gemini, Cursor, Codex, n8n or any MCP client. Connect once; switching AI later doesn’t mean starting again.

about two minutes
step 03

Hand over the task

Ask for the outcome, not the steps. The runtime carries it across your systems and brings back the finished result.

you ask once

works across the systems you already use

  • Email
  • Files and docs
  • Spreadsheets
  • Search and analytics
  • Accounting
  • Rostering and payroll
  • Help desk
  • CRM
  • Calendars
  • Team chat
  • Code
  • Databases
  • Any website
  • All integrations →

Your accounts, your say

It acts on your accounts, only when you ask.

UniversalBench connects your AI assistant to the tools and accounts you already use, so it can read your data and take real actions for you, only on your instruction. It works across many services.

As one example, you connect your own Google Search Console and Google Analytics, and your AI reads your search performance (queries, impressions, clicks, position) and your site traffic to diagnose issues and improve your site. At your instruction it can also submit sitemaps and request indexing.

We act only on accounts you own, only when your AI requests it. We never retain your Google data after returning it to your AI, and we never use it for advertising or to train any AI or ML model.

Full details are in our privacy policy.

Pricing

Start free. Pay only for what runs.

All prices in USD. Pick a monthly plan and hand over the work. Every plan runs on the same runtime, so you only choose how much you need.

Personal

$50/ month

For one person working alone

  • 15 connected apps
  • 5,000 actions a month
  • 250 web searches
  • 5 e-signatures coming soon
  • Password safe
  • Usage and connected apps page
  • Email support

Or start on Free: 2 apps, 1,000 actions a month.

Add to this plan

UB Remote: your AI works on your computer$15 / device
Private database coming soonfrom $15
Phone support$29 / month
Extra apps$5 each
Extra actions5,000 for $60
Choose Personal →

Teams Most popular

$60per person / month

For 2 to 19 people working together

People5= $300 a month

Everything in Personal for each person, plus:

  • 30 connected apps for the team
  • Team dashboard: the team’s work in one place
  • Share passwords with a teammate without showing them
  • UB Remote for every person
  • Shared team database, 5 GB coming soon
  • Choose who can use which app coming soon
  • Hand a task to a teammate’s AI coming soon
  • Shared team memory coming soon
  • Phone support
  • People who only view or approve are free

Add to this plan

Extra apps$5 each
Extra actions5,000 for $60
Setup done for youfrom $500 once
On-site visitfrom $400
Choose Teams →

Enterprise

$150per person / month

For 20 people or more

Everything in Teams, plus the full control layer:

  • Audit page: every action by every person, kept 1 year, exportable coming soon
  • Live activity monitoring with alerts coming soon
  • Approvals before payments and payroll coming soon
  • Staff handover: access cut, work passed to the next person coming soon
  • Company rules checked before the AI acts coming soon
  • Your own private database coming soon
  • Unlimited connected apps
  • 10,000 actions per person
  • Monthly invoice
  • Named contact and on-site visits
  • People who only view or approve are free

Add to this plan

Done-for-you automationsfrom $750 once
A connection to any app, built for youfrom $1,500 once
Extra actions5,000 for $60
Book a call →

Pay yearly and get 2 months free. Actions reset each month and don’t roll over. AI models are billed at a small margin, or free with your own key. Scheduled automations are coming soon.

The control layer: what keeps your company safe while the AI works

What the AI can do grows every month. This part stays the same: a record of everything, rules it must follow, and proof of every job. It’s the same layer we run UniversalBench on ourselves. All included in Enterprise. On Teams, add the audit log and activity monitoring for $25 per person.

Audit log coming soonWho did what, when, from where, with which password, and who approved it. Every person, every action, exportable for your auditor.
Activity monitoring coming soonWatch what every person’s AI is doing as it happens, with alerts for odd hours, new locations or big data exports.
Leash coming soonYour company rules are checked at the moment the AI acts, so it can’t do what you’ve ruled out.
Playbooks coming soonWrite how your company does a job once. The AI follows that playbook every time, so no step gets missed.
Second check coming soonBefore a risky change runs, a second, independent AI reviews it.
Proof of done coming soonEvery finished job comes with a receipt that shows it really happened, not just a “done”.
Staff handover coming soonWhen someone leaves, their access is cut at once and their history passes to whoever takes over the role.
Approvals coming soonPayments, payroll and other sensitive actions wait for a manager’s yes.
Spend guard liveEvery AI call has a cost ceiling checked before it runs. No surprise bills.
Locked-down by design livePasswords are never shown to the AI, and the AI can’t reach your internal network.

What’s an action?

One thing the runtime does for your AI. Planning and safety checks are free.

1 Send an email1 Read your invoices from your accounting app1 Run a calculation1 One page step in a browser1 Save a row to your database

A busy small business uses about 1,700 actions a month. Most people never reach their plan’s limit.

Common questions

The things people ask us first.

Which AIs work with UniversalBench?

Claude, ChatGPT, Gemini, Cursor, and any MCP-compatible client. Switching AI providers does not require re-integrating UniversalBench.

What counts as an action?

One thing the runtime does for your AI: running code, reading or writing a connected app, a database query, a browser step. Planning and safety checks are free.

Can people from another UB account join my team?

Yes, free. Anyone with their own UB account can join your team, and what they use is billed to their own plan, not yours.

Is the 96.5% token reduction real?

Yes, on bulk data tasks. We have measured reductions of 95% to 99.8% across the three live tests above, all run against the current Anthropic flagship and reproducible from published data. Tasks that send small inputs to your AI will not see savings on this scale.

What happens when I use up my plan?

Extra actions are $0.016 each, or 5,000 for $60, or you move up a plan. Actions reset at the start of each month and don’t roll over. Most people use far less than their plan.

Is my data sent to other AI providers?

Only what your AI explicitly asks UniversalBench to send. Credentials you save go into an encrypted vault scoped to your account. Your data is not used for training. Your AI cannot reach your internal network by default.

What if a validation blocks something I want to push?

You see the exact reason. You fix it and try again. The validation is mandatory because that is what makes the safety claim real, but the error is always visible and actionable.

Can I raise the cost ceiling on LLM calls?

Yes. The default ceiling is $0.50 per call, and you can raise it up to $50. The cap stays on, and you control its size.

How do I cancel?

Cancel any time from your account. Your plan stays active until the end of the month you’ve paid for.

Free setup call

Get connected in 30 minutes. I’ll walk you through it.

We connect your AI assistant, run a live task, and make sure everything works. No technical knowledge needed.

Book a free setup call Get API key

30 minutes · video call · 1,000 free actions a month, no card