What is kdeps?

kdeps is a git-native AI appliance builder. You describe an agent in YAML - which model to call, what to validate, what shape the answer takes - and those files are the whole spec. Commit them, and kdeps packages the workflow, its tools, and the model into one self-contained unit you can run as an HTTP API, a terminal REPL, a Docker image, a Kubernetes deployment, a bootable ISO, or a single binary. Same workflow.yaml, no rewrites, no re-config, no framework to import.

Change the YAML and the appliance behaves differently - your git history is the changelog of the agent's behavior. Because it runs open-source models by default, the unit has no per-token cost and no dependency on an external AI service; it works the same on your laptop and inside an air-gapped network. That is data sovereignty: run your own LLM and coding agent on servers you control, in the region you choose, with no prompts or data sent to a third-party AI provider.

The problem it solves

Calling an LLM takes a few lines. Shipping that call as something you can review, version, and run inside your own boundary is not. You end up hand-writing the same glue every time: input validation, retries, ordering between steps, a fixed response schema, a container, a way to run it offline for tests. kdeps is that glue, declared in YAML that lives in your repo. Cloud providers (OpenAI, Anthropic, Groq) still work when you want them; the backend is one line of machine-local config, not part of the repo.

The smallest mental model

a folder with workflow.yaml
        |
        +--  kdeps run ./my-agent/     ->  one-shot pipeline / HTTP API
        |
        +--  kdeps ./my-agent/         ->  interactive REPL, agent calls the workflow as a tool

• A resource is one step - an LLM call, a shell command, a SQL query, an HTTP request. It lives in its own YAML file.
• A workflow is a folder of resources plus a workflow.yaml manifest. Each resource declares what it requires:, and kdeps runs them in that order.
• A mode is how you run the workflow. Workflow mode (kdeps run) executes the steps in a fixed order and returns a structured response - this is what you deploy. Agent mode (kdeps [path]) starts a chat REPL where an LLM decides when to call the workflow.

Every term kdeps uses is defined in the Glossary.

The smallest working example

No YAML at all - just run the binary:

kdeps            # opens an AI chat REPL against a local model, no API key

A minimal workflow is a folder with a manifest and two resources - one that calls the model, one that shapes the response:

# my-agent/workflow.yaml - the manifest
apiVersion: kdeps.io/v1
kind: Workflow
metadata:
  name: my-agent
  targetActionId: response   # which resource produces the final result

# my-agent/resources/llm.yaml - call the model
actionId: llm
chat:
  model: llama3.2:1b         # a local model, downloaded on first run
  prompt: "{{ get('q') }}"   # 'q' comes from the HTTP request body or REPL input

# my-agent/resources/response.yaml - shape the reply
actionId: response
requires: [llm]              # runs after llm
apiResponse:
  success: true
  response:
    answer: get('llm').message.content

kdeps run ./my-agent/     # run it once / serve it
kdeps ./my-agent/         # or load it as a tool in the chat REPL

Which product do you need?

kdeps is a small number of bounded things. Most people need one or two.

You want to...	Product
A deterministic YAML pipeline - API, bot, or file processor	kdeps workflow
A local autonomous agent - tool use, memory, a REPL	kdeps agent
To orchestrate several agents/workflows as one system	kdeps agencies
Just a self-hosted OpenAI-compatible endpoint, no workflow	kdeps LLM server
To ship a tested workflow as Docker / K8s / ISO / a binary	kdeps deploy
To find, install, or publish shared agents by name	kdeps registry (optional)

The registry is the only optional piece - it shares agents by name, it is not a step in building or running your own.

First steps

You want to...	Start here
Build a real HTTP API from YAML	Quickstart
Run an AI agent locally right now	Run locally in 30 seconds
Understand why kdeps works this way	Why kdeps?
See the full picture of every concept	Concepts overview