Features

Build, connect, test and deploy, in one place

Every capability below runs inside your own environment. The agent is a markdown file; the rest is the runtime around it.

Contact sales →

AgentKube is a builder you talk to, a runtime that runs the agents it produces, and the tooling to connect, test and deploy them. Agents are markdown files in a folder you own, so they work with your editor, your git and other hosts.

Build

Describe the process; get an agent you can read.

Agent builder via chat

Describe the agent in plain language, in one prompt or turn by turn. The builder asks only what it cannot infer, plays the flow back, and proposes the file.

Every proposal is validated against the agent file grammar before you see it, so what you create runs.

HOW IT WORKSThe builder works out who starts the conversation, what is collected, what is checked, where it branches, where a person steps in, and how it ends.You create the file from the proposal; the builder never writes to your folder.Change requests become reviewed edits to the same file.

Start from your website

Give the builder your website. Three research agents read it, look for comparable businesses and estimate the impact, then the builder proposes the agents your business needs.

HOW IT WORKSThe brief lists what the business does, who it serves, services, hours and likely customer questions.Research runs with your search key, against your own site.

Steps and branching

An agent is a sequence of steps. Each step has a mode: scripted says exactly one thing, guided follows instructions, autonomous works toward a goal. Steps capture facts, branch on what the customer says, and point at the next step, a handoff or the end.

HOW IT WORKSRouting is decided in code, not by the model, from the branches you wrote.Captured facts are checked before an action runs; a missing fact means another question, not a bad call.

Test before it goes live

A Test tab runs the agent as a customer would, with a Debug view that shows the step, the rule that fired, the facts, every tool call and its result.

HOW IT WORKSTest turns are not counted or remembered.The eval harness grades the builder itself against fixed scenarios with a code grader.

Connect

Your systems, your data, your keys.

Tools and actions

A tool is one file: what it does, its arguments, and how it runs, as a SQL query, an HTTP request or a command. Agents call tools from a step with the facts they captured.

HOW IT WORKSSecrets are never in the file; they are environment variables.The builder proposes tools when a step needs data no tool provides, and asks what system holds it.

Your own MCP

Connect MCP servers once and reuse them across agents. Your business MCP server serves your tool files; the demo server shows the pattern with seeded data.

HOW IT WORKSEach agent lists the tools it may use; it sees nothing else.Servers run inside your network and are protected by your API keys.

Memory

What an agent remembers about people and the business is a folder of plain files, one fact each. The agent proposes; a person accepts or dismisses; the facts that match a conversation are given to the agent.

HOW IT WORKSFiles, or a separate Postgres database when you prefer.Deleting an agent asks whether to keep what it remembers.

Knowledge base

Index your website, folders and uploads. Steps answer from the sources you name and cite the passage they used.

HOW IT WORKSEmbeddings run on your provider; the index lives in your database.Refresh on a schedule or on demand.

Deploy

Where your customers and employees already are.

Teams and Slack

Handoffs reach people on the channels your teams use. Email today; Teams and Slack adapters are on the roadmap.

HOW IT WORKSA packet with the conversation so far, the facts and the reason.

Embed in your apps

The builder itself is embeddable: one element in your own application, your theme, your agents.

HOW IT WORKSAuthenticated with an API key you issue.

Local or your cloud

Docker or Podman on a laptop, your servers, Google Cloud Run, Azure App Service, AWS. Create an environment per place it runs, tag it, and get the bundle and the commands.

HOW IT WORKSEach environment reports its version, its services and whether it is connected.The agent folder is shared through git; the environment tag is not in it.

Operate

People, versions, access.

Handoff and inbox

When a step hands off, the conversation waits in the team’s inbox with full context. A person replies from there; the agent takes the thread back when they are done.

HOW IT WORKSTeams and members come from one file in the folder.

Versions and rollout

Deployments show what each environment runs. Promotion of an agent with everything it needs, gradual rollout and automatic rollback are on the roadmap.

HOW IT WORKSUntil then, environments are promoted through git.

SSO and roles

Sign in with Google or email today; Microsoft Entra ID and Okta per organization through your identity provider. Owners invite members; members build.

HOW IT WORKSThe console closes to everyone else. The widget stays public, behind your keys and domains.

See these features on your own use case

Bring one process. We build, test and deploy it with your team in a technical session.

Contact sales →Run it locally

FAQ

Questions people ask

What does the agent builder do?

The builder interviews you about a process, plays the flow back, and proposes a complete agent file: steps, branches, what to capture, which tools to call and where to hand off to a person. You create the file with one click and test it immediately.

Can an agent call my systems during a conversation?

Yes. A step can run a tool, from your MCP servers or from tool files you define, with the arguments captured earlier in the conversation, and speak the result in the same turn.

Is memory something I can inspect?

Yes. Memory is a folder of plain markdown files per agent. The agent proposes a memory; nothing is saved until a person accepts it. You can read, edit and delete every file.

How do I put an agent on my website?

Paste one script tag. The widget talks to your own runtime, with your organization id and an API key, and only from the domains you allow.