Kilevel
Dashboard

Language / Sprache

KILEVEL © 2026

This page inEnglishDeutsch

For businesses & enterprises
Kios · AI agent operating system

Kios is the system your AI agents are built from, run on and answer to.

State graph agents hold conversations, automations run pipelines and Work agents take on whole jobs in the background. All of them share the same models, knowledge, data and integrations, and answer to the same roles, approvals and budgets.

KiosKiosKiosKiosKiosKiosKiosKiosKiosYour systemsCONVERSATIONSState graph agentsStates, evaluations, toolsand versions on one canvasPIPELINESAutomationsTriggers, nodes, AI stepsand media generationINTEGRATIONSAPI, MCP and webhooksCustom tools, and Kilevelin your AI assistantTHE COREModels and rulesOne set of roles, approvalsand budgets over all of itIN THE BACKGROUNDWork agentsSandboxes, browsers,sub-Works, schedulesKNOWLEDGEKnowledge and SpacesKnowledge bases, Spaces,wikis, repositoriesDATADatabases and applicationsAn agent’s own SQL database,data stores, hosted apps
Six building blocks around one core. Each cell shares its walls with the core and with its neighbours: the same models, rules, data and tools, wherever the work runs.
Three ways to run work

A conversation, a pipeline or a job in the background

Kios gives each kind of work its own builder. Which one fits depends on how the work starts and whether anyone is on the other end, and each of them can start the other two.

IN A CONVERSATIONState graph agentRuns automations and startsWorks mid-conversationIN A PIPELINEAutomationStarts conversationsand Work runsIN THE BACKGROUNDWork agentRuns automations,starts calls and chats
Every corner can start the other two. An agent runs an automation or starts a Work in the middle of a conversation, an automation starts a conversation or a Work, and a Work runs automations and starts calls and chats.
  1. In a conversation

    State graph agents

    A flow is a versioned agent definition: a master prompt, session states, evaluations graded on every turn, and the tools and exits each state may use. The model still speaks for itself; the graph decides where the conversation may go. The same flow runs a phone call, a web chat, WhatsApp, SMS, email or a Work run.

    • States, evaluations and transitions on one canvas
    • Stations released by a check the runtime makes, never by the model grading itself
    • Every save is a version you can compare and reapply
  2. In a pipeline

    Automations

    A graph of nodes started by a trigger: by hand, on a schedule, from your backend, from a hosted webhook, by an agent calling it as a tool, or on a platform event. Its nodes run AI steps, branches and loops, record and API calls and MCP tools, ask a person mid-run, and generate images, speech and video.

    • Six kinds of trigger, one way of running
    • A node-by-node trace of every run: input, output, branch, duration and cost
    • Versions you can compare and restore
  3. In the background

    Work agents

    Give an agent an objective and it plans, acts and checks each step until the job is done, with nobody on the other end. It can work in its own Linux sandbox and a real browser, hand parts of the job to sub-Works, run on a schedule, and stop for a person wherever you said it must.

    • An isolated sandbox VM and a real browser session, which can be recorded
    • Sub-Works with a depth, a parallel count and a budget you cap
    • Approval modes and Human in the Loop stops
Data and knowledge

Agents keep their own records and work from your knowledge

What an agent writes down and what it reads from both live in Kios, under the same rules as everything else.

Data the agent keeps

  • Agent databases

    An agent or a flow can keep its own SQL database, or share one with other flows and agents, and works with it through database tools. You decide whether the AI may only read, what the database is for, its starting tables, its row and storage limits, and whether it is stored in the EU. An explorer lets you browse and edit it, import CSV, run SQL and restore a backup.

  • Workspace applications

    An application holds your records in data stores: customers, appointments, services, feedback, personnel, or stores you define. Agents look a caller up and write back during the conversation, the API keeps a store in step with your own systems, and each application has a hosted page in your logo and colours where your team signs in, with fields hidden by role.

Knowledge it works from

  • Knowledge bases

    Text, PDF, Word, single web pages and whole websites, retrieved passage by passage during the conversation. An item can re-read its source on a schedule, and a shared item is kept once and attached wherever it is needed.

  • Spaces

    Versioned file trees that agents and flows read and write: files and folders, links to skills, tools and knowledge, other Spaces mounted inside, and live views of resources. Agents find things by path, by exact text or by relevance.

  • Wikis

    A company wiki is a Space with sections, pages that link to one another, imports from websites, Notion and Google Drive, and an Ask that answers from the wiki alone, with citations. Edits the AI suggests wait for review.

  • Repository Spaces

    A Git repository connected over HTTPS becomes a Space that stays in sync when you push. The AI may only read it, commit directly, or open a pull request.

  • Memory

    Agents keep durable records and recall them later, scoped to a session, a contact, an agent, a flow, a Space, a user or the whole account.

Integrations and models

Open to your systems, your AI assistant and the models you choose

Everything in Kios can be driven from outside it, and every agent can reach out of it.

  • REST API: Your systems drive Kilevel through a documented REST API, with API keys or OAuth2 client credentials, each limited to the scopes you grant, and an access log of every call.

  • MCP server: AI clients that speak the Model Context Protocol connect to Kilevel’s own MCP server and work through its tools, resources and prompts. Each published agent also has an MCP address of its own.

  • Webhooks: Session webhooks tell your systems when a conversation begins, when it ends and when its analysis is ready. In the other direction, an automation can have its own hosted webhook address that starts it.

  • Tools: Agents act through tools you define: HTTP calls to any endpoint, execution plans that read and write records, other MCP servers, and ready-made actions for Google Sheets, Google Calendar, Slack, Twilio SMS, Notion, Cal.com and Calendly. Keys to other systems are stored once, encrypted.

  • In your AI assistant: A connector of Kilevel’s own lets ChatGPT, Claude, Gemini and Grok read, build and test your agents from inside the assistant, within the permissions you give it. What each assistant may do differs, so the integrations page shows where each one stands.

Governance

Rules the agents cannot talk their way around

The limits live in the system, not in a prompt: who may change what, which actions wait for a person, what a run may spend, and a record of all of it.

  • Roles and sub-users

    Sub-users get one of five roles: Integrator, Manager, Employee, Call Center Agent or Task Agent, and each role opens only what it allows. Other Kilevel accounts can be given access as account managers, without anyone sharing a login.

  • Review before it goes live

    Changes the AI proposes to skills, tools, automations, knowledge and flows wait in a Review inbox. A person approves, rejects, edits or rolls back each one.

  • Approvals inside a run

    A Work run asks before the actions you name, stops at Human in the Loop nodes, and waits a set time for a person before taking the fallback you chose. An automation can park a run until someone answers by link, email or in the app.

  • Budgets

    Each Work run carries limits for spend, tokens, model calls and tool calls, and one more for the whole tree of sub-Works it delegates to, so an agent that starts agents cannot spend past the ceiling.

  • Plan limits

    A plan can cap how many agents, flows, automations, sandboxes, databases, applications and sub-users an account keeps, how many runs go at once, and how many calls and Work runs start each day.

  • Versions and records

    Flows, agents and automations keep their versions, resources an append-only revision history, and the account’s audit log records sign-ins, changes, exports and deletions and who made them, including anyone working in the account from outside it.

Built on Kios

Two products run on the same foundation

For businesses & enterprises

Kitalk · A state-of-the-art AI call center

Kitalk is the AI call center built on Kios, with voice agents for inbound and outbound calls, campaigns, the softphone for your human agents and quality on every conversation.

For personal use, professionals & teams

Ki · Your personal AI executive assistant

Ki is the personal AI executive assistant built on Kios: one assistant per person, in their chats, mail and calendar, keeping track and thinking ahead.

Questions about Kios

What is Kios?

Kios is the name for Kilevel’s foundations: what AI agents are built from and what they run on. It covers state graph agents, automations and Work agents, the databases, applications and knowledge they use, the integrations and models they reach, and the roles, approvals and budgets they answer to.

How do Kios, Kitalk and Ki fit together?

Kitalk and Ki are built on Kios. Kitalk is the AI call center for businesses, Ki the personal assistant for people. When your team builds its own agents, automations and Work agents, it works with Kios directly.

Do we have to build everything by hand?

No. Agents and their state graphs, automations and Work flows can be created and edited with AI assistance, which builds the states, nodes, tools and messages as well as the prompts. Every proposal arrives as a list of changes you accept, reject or edit.

Which AI models do the agents use?

Kios works with frontier models such as OpenAI’s, Claude and KiBrain, and many others, picked for each situation. Kilevel’s own model providers work out of the box, and you can bring your own keys for providers such as OpenAI, Azure OpenAI, Anthropic, Google Gemini, xAI and Mistral.

Can our developers work with Kios from code?

Yes: a REST API with scoped API keys or OAuth2 client credentials, an MCP server for AI clients, session webhooks out and hosted webhooks into automations. The documentation covers each of them.

Can an agent change things on its own?

Only as far as you allow. Changes the AI proposes to flows, automations, skills, tools and knowledge wait for a person in the Review inbox, Work runs ask before the actions you name, and budgets cap what a run and its sub-Works may spend.

Build on Kios

Start with one agent and one flow, or talk to us about running your organisation’s agents on it.