Core concepts
The main ideas in Kavela, including organizations, projects, agents, skills, tools, and evals.
This page defines the main ideas you'll see throughout the docs. You don't need to read it end to end. Skim it now, and come back when a term is unclear.
Organizations
An organization is your account on Kavela. It holds your agents, your projects, and your team. Every organization gets a monthly amount of free credits and the fast Kav-LITE models. See Organizations.
Projects
A project is a file repository inside an organization. You fill it with files, and Kavela's two-layer semantic matcher makes them searchable. Agents pull the right files from a project when a task needs them, which gives them working context. See Projects.
Agents
An agent is the thing you build, publish, and sell. Each agent does one job well, such as reviewing code, drafting outreach, or answering support questions.
An agent is built from two main parts:
- Skills are what it knows.
- Tools are what it can do.
You test an agent with evals before you publish it, and it can draw on your organization's projects for context. On the Kavela website you'll see skills called the agent's brain and tools called its arms.
Skills
A skill is a piece of knowledge written in plain markdown: a runbook, a style guide, the steps your team follows, or the quirks of an API. Kavela stores skills so the agent can find the ones a task needs. Skills are grouped, versioned, and editable. See Skills.
Tools
A tool lets an agent take an action, usually by calling an API. Every agent comes with built-in tools for things like web search, reading files, and creating artifacts. You add more with connections and containers.
Connections
A connection links an agent to an outside account, such as GitHub or Stripe. You can connect your own account as the builder, or have each user connect their own. See Connections.
Containers
A container lets an agent run code and serve a live preview. This is how an agent builds real websites and apps. See Containers.
Evals
An eval is a test case for an agent. You give it an input and check the output. Running evals before you publish shows you which cases pass, which fail, and what to fix. Kavela also scores skills to flag any that don't add something the model couldn't already do on its own.
The context loop
When an agent works on a task, it doesn't put everything it knows into the prompt. It runs a short loop:
Find
The agent searches its skills and project files for what's most relevant to the task.
Load
It reads the full content of what it picked.
Answer
It uses that context to reply, call a tool, or build something.
This is why adding more skills and files makes an agent more capable without making every response slower.