Agents that keep working
Your own agent. It asks first.
Define an agent without code: what it is for, how it speaks, the skills and knowledge it has, a schedule and the tools it may use. Chat with it in real time. When it wants to do something risky, it stops and waits for a person to say yes.
Request access See how it works
Early access, by request. A working name. The windows on this site show sample data.
Connects toGeenius OSany MCP serverthe weba webhook
What it does
Define. Deploy. Approve.
A chat window waits for you. An agent here has its own instructions and memory, can wake on a schedule, and only does the risky things a person has said yes to.
Define
Instructions, persona, skills and knowledge, as data. Every save is a version.
AgentsConnect
MCP connectors, the web and a webhook. A tool is risky unless you mark it read-only.
ToolsWake
A schedule from every minute to every day, and a digest of the last run.
SchedulesApprove
A risky call is suspended until a person decides in the approvals inbox.
Approvals
Agents as data
No code, every save a version.
An agent is a definition: a description, standing instructions, a persona, skills and knowledge as text, a model tier, connectors, tools, a schedule and limits. It is validated when you save, stored as an immutable version and obeyed by the runtime. Deploy a version, roll back to another, and the agent follows.
- Skills and knowledge are text in the prompt
- Templates to start from: inbox triage, research digest, a Geenius OS project reporter
- Folders, tags and favourites for a library
The risk rule
One rule decides what asks first.
Reading the web is safe. Every connector tool is risky unless you listed it as read-only. A webhook is always risky. Anything the rule does not know is risky. A risky call is suspended inside the agent and appears in the approvals inbox; the agent waits until a person approves or denies it.
- The approval gate reads only this rule
- A decision is recorded and the agent is told
- No one can approve through the AI door
Schedules
It wakes up and leaves a digest.
Give an agent a schedule and a prompt, and it wakes by itself: read the pages you follow, report what is open in your projects. The last result is kept as a digest you can read, and your own AI can read it too through Geenius OS.
- From every minute on Max to every hour on Pro
- A hard stop on tool calls a day
- Logs and costs per run
Your model
Your own key, or built-in credits.
On Pro an agent runs on the model key you add; it is sealed and never shown again, and the deployment's keys are never used for your workspace. On Premium and Max agents run on the built-in AI and a month's credits are counted per step, by model tier. A tool-call limit a day is a hard stop, not a bill.
- Pro: 1 agent, your own key
- Premium: 3 agents, 1,000 credits
- Max: 10 agents, 3,000 credits
A door for your AI
Your AI can watch and talk. Not decide.
Connect Claude, ChatGPT or Grok to Geenius OS once. Your AI can list your agents, send one a message, read a digest, look at the approvals queue and check usage. It has no approve action: otherwise an AI could approve its own webhook, which is what the gate exists to prevent.
- One tool, six actions
- On every plan
- Connector secrets are never readable
In numbers
Facts you can check, not testimonials.
Agent host is in early access and has no public customers yet. These are things you can look up in the app and the docs today.
- 1rule decides what is risky, and the gate reads only that
- 3built-in templates: inbox, research digest, OS project
- 6actions on the AI door, none of them approves
- 10agents on Max
Honest labels
What works today.
We say what is live in early access, what depends on your setup, and what is not offered.
- Pre-betaAgents, chat, connectors, approvals, schedules and versions
Built and in the app now. Early access: people request access.
- Pre-betaBuilt-in AI on Premium and Max
The model is called through the hub with the deployment's keys. On a development deployment without keys an offline stand-in answers; there is no such stand-in for a customer.
- Pre-betaLimits and credit costs
The plan limits and the cost of a model step per tier are the current values and may change before plans are sold.
- Pre-betaThe name
Agent host is a working name.
- Not offeredYour own code
An agent is data: there is no code to run.
- Not offeredApproving through your AI
Only a person of the workspace decides an approval, in the app.
What agents do here
For people who want help that asks first.
Small business owners
It labels, summarises and drafts replies. It never sends.
See the setupPeople who follow things
It reads the pages you name and tells you what changed.
See the setupGeenius OS users
Connect Geenius OS and have it report what is open and blocked.
See the setupPrompt Verse authors
Author the definition in Prompt Verse and link it to a hosted agent.
See the setup
Pricing
Three plans. No free tier.
Per month, in euros, VAT included. Included in the matching Geenius OS plan. Payments are in test mode during early access.
- €9Pro: 1 agent, your own model key
- €19Premium: 3 agents, 1,000 credits
- €49Max: 10 agents, 3,000 credits
From the blog
Notes on agents that ask first.
An agent is a definition, not a program
Why Agent host defines an agent as data: description, instructions, skills, knowledge and tools, saved as versions you can roll back.
What counts as risky, and who decides
The one rule Agent host uses to decide which tool calls an agent may make alone and which wait for a person, and how to mark tools safe.
A digest while you sleep
How to set up an Agent host agent that wakes on a schedule, reads the pages you follow and leaves a short digest you read in the morning.
Questions
Good questions.
What is Agent host?
A place to create and run your own agents without code: persona and instructions, skills, knowledge, schedules, tools through MCP connectors, a real-time chat and an inbox for approvals of risky actions. It is an app of Geenius OS and also works on its own.
Is Agent host the final name?
No. It is a working name and can change.
Is it ready to use?
It is in early access: people request access and payments are in test mode. What this site describes is built. Request access.
Do I write code?
No. An agent is a definition: text, settings and tools. There is no code to run.
Who makes it?
AVS Media, a small company in Nijmegen, the Netherlands, as one of the apps of Geenius OS.
Do I need Geenius OS?
No. Agent host has its own account and plans. A Geenius OS plan includes the matching plan, and Geenius OS is the first connector an agent can use.
Give it a job.
Agent host lets people in step by step. Ask for a place, or sign in with Geenius OS.