POMEX AI PLATFORM

Built for enterprise AI agents.

The quiet infrastructure behind agents that do real work.

agent.py

from pomex import Pomex
px = Pomex()   # one key for models, sandboxes and workflows
# Think: any frontier model, through the gateway
reply = px.chat(model="openai/gpt-sol-latest", messages=msgs)
# Act: run the agent's code in an isolated microVM
sbx = px.sandbox.create(size="standard")
sbx.commands.run("python analyze.py")
sbx.kill()
# Deliver: brief in, reviewed video out
run = px.workflows.run("video-production", brief="60s Q3 product recap")

$ python agent.py

✓ reply     openai/gpt-sol-latest · 200 · 0.41 s

✓ sandbox   sbx_7f3a ready · 2 vCPU / 4 GiB

● workflow  run_2048 · waiting on storyboard approval

One key.

pmx_live_••••••••3f9a

No more juggling provider keys, quotas and invoices. One Pomex key for every model, sandbox and workflow, on one bill.

01 / PRODUCTS

Think. Act. Deliver.
Governed end to end.

THINKAvailable

AI Gateway

One API and one key for frontier models. Routing and failover across providers, per-team keys and budgets, unified usage and billing.

ClaudeGPTGeminiDeepSeekQwen+ more

ACTAvailable

Agent Sandbox

A Firecracker microVM per agent, up in seconds. Run generated code, process files, analyze data. Long-lived keys never enter the box. E2B-compatible.

● sbx_7f3a python 2 vCPU / 4 GiB running

● sbx_91c0 python 4 vCPU / 8 GiB running

● sbx_02be python 2 vCPU / 4 GiB destroyed

DELIVERDesign partners

Agentic Workflows

Vertical, multi-agent pipelines that produce finished work — for media, sales and documents. Humans approve at every gate.

brief · script · storyboard · render · review

02 / WORKFLOWS

Agent teams that hand you finished work.

Each workflow is a crew of specialist agents with a producer's checklist — built on the gateway and the sandbox, paused wherever a person should decide.

Start a pilot

MEDIADesign partners

Video production

Brief in, reviewed cut out. A crew of agents scripts, storyboards, generates every shot and assembles the edit, and stops for your approval twice.

THE CREW

ProducerWriterCriticDirectorEditor

shots

0:60

final cut

approval gates

RUNS ON

AI Gatewayscript, image and video models

Agent Sandboxone microVM per render job

See how a run works →

run_2048 · Q3 product recapRendering · shot 5 / 8

00:23 / 01:00SHOT 05 · DOLLY INQ3 is when the agents started shipping.

✓ Storyboard approved○ Final review · after renderExport · MP4 + SRT

SALESComing soon

Sales outreach

Agents research each account, write a sequence from real signals and keep the CRM current. A rep approves every message before it sends.

Accounts 24Drafts 6Sent 18

Northwind Logistics

signal · hiring 4 platform engineers

researched

Helio Health

signal · raised Series B

draft ready

Quanta Retail

signal · new CTO started

awaiting rep

To: VP Engineering, Quanta Retail · step 1 of 3

Congrats on the new role, and on the agent roadmap

Approve & sendEditrep approval required

Researcher · Writer · CRM sync

DOCSComing soon

Document production

Proposals, RFP responses and reports drafted from your knowledge base. Every answer cites its source, and a reviewer signs off.

Security questionnaire · Northwind RFP112 / 120

Q14 Is customer data encrypted at rest?

security-policy.pdf · p.4

Q15 How are production keys rotated?

runbook-keys.md

Q16 Which regions can data be stored in?

needs reviewer

Reviewer · on Q16

Confirm the EU region option with legal before we send.

Retriever · Drafter · Checker

Interfaces shown are product concepts with sample data.

VIDEO PRODUCTION · ANATOMY OF A RUN

Five stages. Two approval gates. One invoice.

Brief

Goal, audience, length, brand kit and any source footage.

agent · intake
runs on · gateway

Script

Drafted, checked against the brief and brand rules, revised.

agent · writer + critic
runs on · gateway

Storyboard approval gate

Shot list with reference frames. You approve before anything renders.

agent · director
runs on · gateway (image models)

Render & assemble

Shots generated in parallel, then cut together with captions and music.

agent · editor
runs on · sandbox + gateway (video models)

Review approval gate

Comment on the cut, request changes, or approve for export.

you
output · MP4 + SRT captions

You choose the models

Every stage calls models through your gateway, under your organization's allowlist.

Cost per run, not per guess

Model and compute spend for each run, broken down by stage, on the same bill.

Isolated by design

Heavy steps run in their own microVM sandbox and are wiped afterwards.

03 / ENTERPRISE

Let agents in. Keep everything else out.

Isolation, scoped credentials and spend controls are defaults on Pomex, not add-ons — so your first agent and your hundredth pass the same review.

Talk to sales Security overview

ARCHITECTURE

Where an agent's request goes — and where it can't.

YOUR AGENT

Agent / app

holds a Pomex key
scoped per team

POMEX

Policy layerallowlists · rate limits · budgets · billing

AI Gateway

provider keys stay here

Sandbox microVM

scoped token only
✕ internal network
✕ cloud metadata

UPSTREAM

Model providers

Anthropic · OpenAI
Google · open models

CONTROLS

On by default.

Control	GATEWAY	SANDBOX	WORKFLOWS
Model allowlists	Per organization	Inherited	Inherited
Rate limits & budgets	Per org and key	Per sandbox	Per run
Credential exposure	Provider keys server-side	Scoped, revocable token	Scoped, revocable token
Isolation	—	Own kernel per microVM	Per stage
Data lifecycle	Prompts not stored by default	Wiped on destroy	Kept until you delete the run
Human approval	—	—	Configurable gates
Billing	One invoice — model + compute, by team, key, sandbox and run

Compliance

Security whitepaper, architecture review and questionnaire responses available under NDA.

Identity

Team-scoped API keys and role-based console access. SAML SSO on enterprise plans.

Regions

Pin model traffic to specific provider regions and choose where sandboxes run.

Procurement

Annual agreements, monthly invoicing and custom terms through your account team.

FAQ

Questions security teams ask first.

Can an agent in a sandbox reach our internal network?

No. Internal networks and cloud metadata endpoints are blocked by default. Each sandbox is its own Firecracker microVM.

Do our model provider keys ever reach the agent?

No. Inside a sandbox the agent gets a scoped token that you can revoke. Long-lived keys never enter it.

How do we cap spend?

Set budgets and rate limits per organization and key at the gateway. Every cost — model and compute — appears on one bill.

We already use E2B. How hard is it to move?

The sandbox is E2B-compatible. Point two environment variables at Pomex and keep your code.

Talk to our team.

Tell us what your agents need to do. We'll come back with an architecture and a pilot plan.

Sales: sales@pomex.ai
General: info@pomex.ai

Built for enterprise AI agents.

One key.

Think. Act. Deliver.Governed end to end.

AI Gateway

Agent Sandbox

Agentic Workflows

Agent teams that hand you finished work.

Video production

Sales outreach

Document production

Five stages. Two approval gates. One invoice.

You choose the models

Cost per run, not per guess

Isolated by design

Let agents in. Keep everything else out.

Where an agent's request goes — and where it can't.

On by default.

Compliance

Identity

Regions

Procurement

Questions security teams ask first.

Talk to our team.

Think. Act. Deliver.
Governed end to end.