Skip to content

Kampong Agents

Build an agent workflow once. Keep it as a file you can read, diff, and review.

Kampong Agents is a builder for agent workflows. You design one on a visual canvas, or hand-write its YAML in your editor, and either way there's exactly one definition: an AgentSpec that both views share. Locally that's a file you can commit; in the hosted product it's managed for you and you can import or export it as YAML whenever you like. When the agent is ready, you can eject it to a standalone TypeScript project that no longer depends on this tool at all.

It runs locally. Nothing phones home, secrets stay in your environment, and a run can execute fully offline against a local model.

The one idea worth knowing first

The canvas and the YAML are not two things kept in sync. They are the same thing. Every edit on the canvas rewrites the YAML, and every edit to the YAML re-renders the canvas.

flowchart LR
  canvas["Visual canvas"]
  spec["agent.yaml<br/>(the spec, as a file)"]
  editor["Your editor<br/>(Cursor, Claude Code, vim)"]
  export["Standalone<br/>TypeScript project"]

  canvas <-->|"read + write"| spec
  editor <-->|"read + write"| spec
  spec -->|"kampong export"| export

You can start on the canvas and finish in your editor, or the other way around, and the file is always the truth. It's plain YAML, so it belongs in git next to your code.

A 60-second taste

Here's a complete agent. It greets the user and answers one question:

# yaml-language-server: $schema=./agent-spec.v1.0.schema.json
version: "1.0"
agent:
  id: hello_agent
  name: "Hello Agent"
  role: "Friendly assistant"
  goal: "Greet the user warmly and answer one question in a concise, helpful tone."
  model:
    provider: openrouter
    name: liquid/lfm-2.5-2.6b:free
    api_key: ${OPENROUTER_API_KEY}
  workflow:
    - step: greet
      action: generate_text
      inputs: [input]

Run it from the command line:

kampong run agent.yaml --input "Hi, what can you do?"

Or open it on the canvas and press Run:

kampong dev

Both execute the same agent, against the same file. That's the whole idea.

Where to go next

  • Getting started installs the tool and runs your first agent in a couple of minutes.
  • Tutorial: your first agent builds up from hello-world, one concept per step: models, tools, workflows, guardrails.
  • Concepts covers the design ideas: canvas/YAML duality, local-first execution, one-way export.
  • Reference documents every CLI command and every field of the AgentSpec schema.

What it does not do yet

Kampong Agents is early. Today it builds one agent at a time, with no multi-agent org charts. It runs and exports that agent, works offline, and treats hand-editing as a first-class workflow. A workflow can run on a schedule or a webhook, loop over a list, branch, and call first-party connectors (see Scheduled data pipelines). The canvas edits all of this, including steps inside a loop or a branch arm, and an assistant can build or change a workflow from a description (see the workflow assistant). There is also a hosted app, the same canvas with sign-in, workspaces, and server-side provider keys. It is built but not yet deployed. It runs version 1.1 specs, except schedule and poll triggers. Enterprise governance and multi-agent orchestration are on the roadmap, not in the box. See Design decisions for the reasoning behind the current scope.