Agent Runtime · private beta

Run AI agents in production, sovereign by design.

One API call runs your agent. Each conversation gets its own VM in the EU.

Request beta accessHow it works

Your model key. Per-second pricing. No conversations stored.

app.py
>>> agent.invoke("sha256 of orkestr-sdk, first 12 chars?")
run.acceptednew session, its own VM
tool.callrun_shell
tool.resultok
"2e95fa121c8b"
run.completed2 steps
POST api.orkestr.eu/v1/agents/support-bot/invoke
beta
1-minute tour

Agent Runtime in 60 seconds

Everything an agent needs to go live.

Managed for you, in the EU.

Its own VM

Each conversation runs isolated, with Python, shell, files and the web built in.

Your tools

Plug in MCP servers and APIs. The keys stay with orkestr, never the agent.

Your model

Bring a Mistral, OpenAI or self-hosted model. Tokens bill to your key.

Memory per user

Optional facts about each of your users, recalled in every new conversation.

Versions

Publish a version, point prod at it, roll back in one move.

Approvals

Risky tools pause the run until you say yes.

Build it your way.

Call it from your code, or let your coding assistant build it.

  • Python and TypeScript SDKs, or plain HTTP
  • Wait for the answer or stream each step
  • Invoke-only keys for your app servers
from orkestr import Agent

agent = Agent.get("support-bot")

result = agent.invoke("Where is my order #4411?",
                      end_user_id="cust-42")
print(result.output)

# Same conversation, same VM:
agent.invoke("Can you change the address?",
             session_id=result.session_id)

Sovereign by design

EU hardware, and less data to protect.

EU operator, EU hardware

No US parent, no Cloud Act reach.

EU end to end

Pair it with an EU model and nothing leaves.

GDPR in one call

Erase one of your users everywhere, instantly.

We keep

  • Agent settings and versions
  • Run records: status, steps, tool names
  • Token counts per end user
  • Memory, only if you turn it on

We never keep

  • The conversation itself
  • Your model key inside a VM
  • Your tool keys inside a VM
  • Live events after one hour

The life of a conversation.

One conversation, one VM. Click a stage.

Parked. Paused to a snapshot: storage only, back in under a second.

< 1 s

to resume a parked conversation

1 VM

per conversation, never shared

0

conversations in our database

Pay for seconds, not calls.

Same rates as sandboxes. Tokens bill to your own model key.

vCPU

€0.045

/ hour

RAM

€0.015

/ GiB-hour

Storage

€0.10

/ GB-month

A default session is 1 vCPU and 1 GB.

Common questions

What is the Agent Runtime?
A managed place to run AI agents you call from your product. You define the agent; we run each conversation in its own EU VM.
Do you store our users' conversations?
No. A conversation lives in its VM and is deleted with it. We keep run records and token counts, never the messages.
Which models can I use?
Mistral, OpenAI, or any OpenAI-compatible endpoint, including your own. Pick an EU model and nothing leaves the EU.
Where does it run?
On hardware we operate in Germany, Finland and France, run by an EU company with no US parent.
How is it priced?
Per second of compute while a session runs, plus storage while it is parked. No per-call fee; tokens bill to your key.
Can my coding assistant build the agent?
Yes. Paste one line into Claude Code, Cursor or similar, and it builds, tests and publishes the agent for you.
How do I get access?
It is in private beta. Ask for access and we turn it on for your account.

Put your agent in production, in the EU.

Private beta. Tell us what you are building.

Request beta access