Skip to content
New

Ozmio Agents 2.0 is here

2.0 Multi-agent handoffs are live

Agents that actually ship work.

Ozmio is the platform to build, orchestrate and observe AI agents that use your tools, follow your policies and hand off to humans when it matters.

$ npx ozmio init support-agent
  • Free up to 1,000 runs
  • SOC 2 Type II
  • EU & US data residency

ops-assistant · run #4,812

Idle

Task

Live trace

Tokens

0

Cost

$0.000

Elapsed (s)

0

Tool calls

0

Progressozmio-large
  • Guardrails3 active
  • ApprovalsRefunds > $200
  • PIIRedacted

Trusted by 2,400+ teams running agents in production

  • Halvard
  • quorbit
  • Lumenfield
  • Northfield
  • Taskwell
  • vantapoint
  • Orchidline
  • Kestrel

01 Platform

Everything an agent needs to do the job

Tools, memory, policies and evals in one place — so agents move from impressive demos to dependable teammates.

Tools & integrations

Connect 180+ apps or any REST/MCP endpoint. Every call is typed, permissioned and logged.

180+ tools
  • Helpdeskconnected
  • Warehouseconnected
  • Emailconnected
  • Billingconnected
  • Web searchread-only
  • Custom API+ add tool

Long-term memory

Agents remember customers, decisions and preferences across runs.

  • acct.halvard.plan = "Scale"2m
  • pref.tone = "concise, friendly"1d
  • policy.refund_limit = 2006d

Guardrails & approvals

Write policies in plain English. Risky actions pause for a human.

  • Redact PII in logs
  • Approve refunds > $200
  • Allow external email

Evals on every release

+6.4%

Regression-test agents against real conversations before they ship.

  • Task success94.2%
  • Policy adherence99.1%
  • Hallucination rate0.8%

Observability

p50 1.8s

Trace every step, token and cent — and replay any run.

Deploy where work happens

Ship one agent to every channel without rebuilding it.

  • Website widget
  • Team chat
  • Email
  • Voice
  • API & webhooks

02 Orchestrate

Multi-agent flows, with a human in the loop

Chain specialist agents, branch in parallel and pause for approval. Drag the nodes — it’s a real canvas.

flows/refund-requests.yaml

Drag nodes or press Run

idle

New ticket

trigger · helpdesk

idle

Planner agent

classify · route

idle

Research agent

kb.retrieve · history

idle

Billing lookup

tool · billing.lookup

idle

Human approval

if refund > $200

idle

Reply & close

helpdesk.reply

  • Handoffs & parallel branches

    Specialist agents share context and memory, so nothing is lost between steps.

  • Human-in-the-loop

    Approvals land in team chat or email with full context and one-click decisions.

  • Versioned as code

    Flows live in YAML or TypeScript, reviewed in pull requests and rolled back in seconds.

4.2B

tokens routed every month

38ms

median orchestration overhead

99.99%

API uptime, last 12 months

71%

of tickets resolved without a human

03 Playground

Talk to an agent before you build one

Pick a model, tune the temperature and switch tools on or off. Then ask about a refund, a meeting or your best accounts.

0.4
precisecreative
Tools
  • Billing
  • Calendar
  • CRM
  • Docs search

System prompt

You are Ops Assistant for Northfield. Be concise. Never issue refunds above $200 without approval.

Ops Assistant

online
Hi! I’m connected to billing, calendar and CRM. What should we get done?
Which accounts are ready to expand?
crm.search()12 accounts match “expansion ready”Twelve accounts look expansion-ready. The top three — Halvard, Quorbit and Lumenfield — each grew seat usage by more than 40% this quarter. Want me to draft outreach?

04 Developers

Ten lines to your first production agent

Typed SDKs for TypeScript and Python, a REST API for everything else, and local dev that mirrors production.

  • Streaming, retries and tool schemas built in
  • Bring your own model or use Ozmio’s routing
  • OpenTelemetry traces out of the box
import { Agent, tools } from "@ozmio/sdk"; const agent = new Agent({ name: "ops-assistant", model: "ozmio-large", tools: [tools.helpdesk, tools.billing, tools.crm], memory: { scope: "account" }, policies: ["approve refunds over $200"], }); const run = await agent.run("Triage the overnight inbox"); // → streamed steps, tool calls and a final report console.log(run.summary);

05 Customers

Teams that put agents on payroll

“We replaced a patchwork of scripts with three Ozmio agents. Support now clears the overnight queue before anyone logs on — and every refund still gets a human sign-off.”

Maya RahmanVP Customer Operations, Halvard −63%time to first reply 9 wksto full rollout

“The trace view is the reason we trust it. We can see exactly why an agent did what it did.”

Daniel KovačStaff Engineer, Quorbit

“Evals caught a regression before it ever reached a customer. That alone paid for the year.”

Amara OkaforHead of AI, Lumenfield

“Finance closes the month two days faster. The reconciliation agent does the boring 90%.”

Jonas LindqvistController, Northfield

“Policies in plain English meant legal could review them. That unlocked production in a week.”

Sofia PereiraCOO, Taskwell

“We prototyped in the playground on Monday and shipped to our help center on Friday.”

Ravi ThakurProduct Lead, Vantapoint

06 Pricing

Start free. Scale when it works.

Simple plans with usage included. No per-seat surprises.

Save 20%

Starter

For builders exploring agents.

$0/ month

Start for free
  • 1,000 agent runs / month
  • 3 agents, 20 tools
  • Playground & traces (7 days)
  • Community support

Scale

For high-volume, regulated workloads.

$159/ month, billed yearly

Talk to sales
  • 250,000 agent runs / month
  • SSO, SCIM & audit logs
  • EU / US data residency
  • Dedicated success engineer
  • 99.99% uptime SLA

Enterprise

Private deployment, custom models, volume pricing and a security review.

Contact sales

07 FAQ

Questions, answered

Still deciding?

Book 30 minutes with an engineer. We’ll build a working agent on your data, live.

Book a call
What counts as an agent run?

One run is a single task from start to finish, however many steps or tool calls it takes. Playground runs are free and don’t count toward your limit.

Can I use my own models?

Yes. Bring keys for any hosted model or point Ozmio at your own inference endpoint. Routing, evals and traces work the same either way.

How do approvals work?

Any step can require a human. The run pauses, a reviewer gets the full context in team chat or email, and the agent continues the moment they approve or edit.

Is my data used to train models?

Never. Your data stays in your chosen region, is encrypted at rest and in transit, and is deleted on request.

Do you support SSO and audit logs?

SSO (SAML/OIDC), SCIM provisioning and exportable audit logs are included on the Scale and Enterprise plans.

Can I cancel any time?

Yes — monthly plans cancel at the end of the period, and yearly plans can be refunded pro-rata within the first 30 days.

Put your first agent to work today.

Free up to 1,000 runs a month. Set up in minutes, scale when it pays for itself.