Run AI workflows locally. Or deploy them anywhere.
Install kdeps, run `kdeps`, get an AI agent - no API key needed with Ollama or llamafile. Build your workflow in YAML. Deploy as Docker, Kubernetes, or a single binary when you're ready. Proud member of the NVIDIA Inception program.
$ kdeps
kdeps v2.x.x | Local agent mode
Model: llama3.2 (Ollama) | Type /help for commands
> write a Go function that parses a CSV file
Sure. Here's an idiomatic Go CSV parser...
func ParseCSV(r io.Reader) ([][]string, error) {
reader := csv.NewReader(r)
return reader.ReadAll()
}
> /model claude-opus-4-8
Switched to claude-opus-4-8 (Anthropic)
>getting started
Three steps from idea to running AI API.
Declare resources -- chat, HTTP, Python, SQL, shell. Wire them with requires:. No glue code.
kdeps run workflow.yaml starts the API server. kdeps serve starts the autonomous agent loop.
POST to your route, get back structured JSON. Export as Docker, Kubernetes, ISO, or a single binary.
platform agnostic
Switch backends in config. No code changes, no lock-in.
run modes
Workflows, agents, and agencies — all from the same YAML.
Resources run in DAG order defined by requires:. Every request takes the same path. Predictable, auditable, ships to production.
Run kdeps for an instant AI REPL - fully offline with llamafile or Ollama, no API key needed. Load workflows as tools: the LLM decides what to call and when.
One agent calls another via the agent: resource type. Compose agents like functions — each runs independently, results flow back.
why kdeps
No Python scripts, no wiring, no boilerplate.
| Traditional approach | kdeps |
|---|---|
| Python script + Flask + OpenAI SDK + retry logic | One workflow.yaml file |
| Manual dependency wiring | requires: in YAML |
| Write error handling by hand | onError: block (continue / retry / fail) |
| Write Dockerfile + CI pipeline for deployment | kdeps bundle build --tag then docker push |
| Glue code between services | Resources pass data via output() |
| Manual polling loop for bots | input.sources: [bot] in workflow config |
| Chaining agents by hand | agent: resource — one agent calls another declaratively |
registry
Install pre-built packages from the registry. Publish your own.
Reusable capability extensions. Install with one command, invoke with component: and typed inputs.
Complete DAG pipelines packaged as .kdeps archives. Drop them into an agency or run them standalone.
Multi-agent orchestration bundles packaged as .kagency archives. One entry point, many agents.
examples
Real patterns from the examples directory. Every one is a working workflow.
POST a JSON body, run a DAG pipeline, get structured JSON back. The default pattern for workflow mode.
workflow modePoll for messages, run a multi-step pipeline, reply with LLM responses. Two resources: llm and reply.
workflow modeRead from stdin, call an LLM, write to stdout. One-shot. Perfect for cron jobs and CI pipelines.
workflow modeIndex documents locally, search with keywords, feed results into an LLM prompt. Fully on-prem.
workflow modeOrchestrate multiple agents. One summarises, another translates — each an independent workflow.
multi-agentbook
Everything from first agent to production deployment, in one place.
Build & Deploy Autonomous AI Agents and Agencies in YAML
Your AI prototype works. Now ship it. Hands-on guide to deterministic pipelines, multi-agent orchestration, error handling, and vendor-agnostic deployment — the production challenges most AI frameworks leave to you.
Read the book ->