First steps

3 min read

These five steps will get your first agent ready to work. The rest of the configuration can wait: add capabilities when you need them.

If your organization does not yet have a connected machine or a usable model, the dashboard shows a full-screen assistant that asks for them, and does not let you see the rest until they are in place. If it cannot check, it lets you through to the dashboard. These steps follow the same order.

Before you start

Confirm your email address. Until then, you can explore the dashboard, but you cannot create organizations, start agents, connect channels, or invite your team. See Your account if you need help.

Set up your organization

Set its name and time zone. If other people will administer it, invite them now. The organization contains your agents' shared settings and resources.

Set up the organization · Manage the team

Connect a machine

Agents run on a machine in your infrastructure: your own computer or a server. The dashboard offers the install command for Windows, macOS and Linux; the servers guide covers Windows and Linux in detail. To start, your computer is enough and costs nothing, but agents only work while it is on; whenever you like, you move them to a server. Check that it is online before assigning an agent to it.

Connect a server

Add an AI model

Connect your subscription or an API key from a cloud provider, or a local engine such as Ollama or vLLM. In the assistant, Test checks the connection and only switches it on if it works. New agents use the organization's default credential until you assign a different one.

Set up models and credentials

Create your first agent

Choose a name and the server on which it will run. You can then define its instructions, knowledge, tools, and access.

Create an agent

Connect it to a channel

Connect a channel if it needs to receive messages or events from people and other systems. Skip this step if it will only perform internal tasks.

See the available integrations

Set it up with an external agent

If you prefer to delegate this setup to Claude Code, Codex, or another coding agent, send it this single message:

Fetch https://api.kujira.so/agent-setup/prompt.md

The agent will download the official instructions for connecting to Kujira and ask you for a service token with the permissions it needs. From there, it can guide you, complete the configuration through the API, or import an agent you already have in another tool, such as a Claude Code or Codex project, or a custom GPT.

Grant only the permissions that are needed

The token determines what the external agent can view or change. Start with read permissions and expand them only if you want it to make changes. Do not paste the token into the conversation: save it in the KUJIRA_API_TOKEN environment variable and restart the agent so it can read it.

After your first agent

Define how it works

Configure its identity, permissions, limits, and resources for the work it will take on.

Provide context

Add memories, skills, and tools to give it the context and access it needs.

Supervise its activity

Decide when it must ask for approval and review the actions that require a decision.

You can work in stages

You can add more servers, models, agents, and channels when your use case requires them.