KDeps

Skip to content

Skills for AI agents are hereRun AI workflows locally. Or deploy them anywhere.<br>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.

Try locally<br>Get Started<br>Why kdeps?<br>Registry

local agentworkflow.yamlresources/fetch.yamlresources/summarize.yaml

$ kdeps

kdeps v2.x.x | Local agent mode<br>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) {<br>reader := csv.NewReader(r)<br>return reader.ReadAll()

> /model claude-opus-4-8<br>Switched to claude-opus-4-8 (Anthropic)

>$export KDEPS_API_AUTH_TOKEN=dev-token<br>$kdeps run workflow.yaml<br>Listening on :16395

$curl -s -X POST localhost:16395/summarize \<br>-H "Authorization: Bearer $KDEPS_API_AUTH_TOKEN" \<br>-H "Content-Type: application/json" \<br>-d '{"url": "https://example.com"}'

{"success": true, "data": {"response": "Example.com is used for illustrative examples in documentation."}}

Local AI agent<br>Run `kdeps` and you are in an AI REPL. Use Ollama or llamafile for a fully offline, private coding agent - no API key, no cloud dependency.

Portable workflows<br>Write a workflow.yaml once. Run it locally with `kdeps run`. Deploy it as Docker, Kubernetes, or a binary when you are ready - same file, no rewrites.

Agencies<br>Multi-agent orchestration. One agent calls another declaratively via the `agent:` resource type. Compose agents like functions.

Any backend, no lock-in<br>llamafile and Ollama work out of the box - no server install, no API key. Or use OpenAI, Anthropic, Groq, and any OpenAI-compatible endpoint. Switch backends in config without touching workflow files.

Component registry<br>Install pre-built scraper, search, browser, and embedding components. Compose them into your workflow with one line.

Deploy anywhere<br>Export as a Docker image, Kubernetes manifests, bootable ISO, or a self-contained single binary. Apache 2.0 license. Standard YAML in a git repo - no proprietary format.

Validate and doctor<br>kdeps validate checks schema, dependencies, and expressions. kdeps doctor diagnoses environment issues before you hit run.

Hot reload dev mode<br>kdeps run --dev watches your files and reloads on change. Iterate fast without restarting the server.

Skills for AI agents<br>A coding-agent skill teaches Claude Code, Cursor, and other agents how to scaffold kdeps workflows, components, and agencies automatically.

getting started<br>How it works<br>Three steps from idea to running AI API.<br>01<br>Write YAML<br>Declare resources -- chat, HTTP, Python, SQL, shell. Wire them with requires:. No glue code.

02<br>Run kdeps<br>kdeps run workflow.yaml starts the API server. kdeps serve starts the autonomous agent loop.

03<br>Call the API<br>POST to your route, get back structured JSON. Export as Docker, Kubernetes, ISO, or a single binary.

platform agnostic<br>Works with any LLM backend<br>Switch backends in config. No code changes, no lock-in.<br>LLM BackendsOllamalocal<br>OpenAIcloud<br>Anthropiccloud<br>Groqcloud<br>LM Studiolocal

Deploy TargetsDockerdocker image<br>Kuberneteshelm chart<br>Binarysingle file<br>ISObootable image<br>WebServerhttp server

run modes<br>Three ways to run<br>Workflows, agents, and agencies — all from the same YAML.<br>workflow<br>Deterministic pipelines<br>Resources run in DAG order defined by requires:. Every request takes the same path. Predictable, auditable, ships to production.<br>RequestvalidatellmapiResponse

$ kdeps run workflow.yaml<br>Learn more -><br>agent<br>Local AI agent<br>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.<br>PromptLLMToolsResponse decidesresultanswer

$ kdeps<br>Learn more -><br>agency<br>Multi-agent orchestration<br>One agent calls another via the agent: resource type. Compose agents like functions — each runs independently, results flow back.<br>CallerSummariserTranslatorResponse agent:result

$ kdeps run agency.yaml<br>Learn more ->

why kdeps<br>YAML replaces glue code<br>No Python scripts, no wiring, no boilerplate.<br>Traditional approachkdepsPython script + Flask + OpenAI SDK + retry logicOne workflow.yaml fileManual dependency wiringrequires: in YAMLWrite error handling by handonError: block (continue / retry / fail)Write Dockerfile + CI pipeline for deploymentkdeps bundle build --tag then docker pushGlue code between servicesResources pass data via output()Manual polling loop for botsinput.sources: [bot] in workflow configChaining agents by handagent: resource — one agent calls another declaratively

registry<br>Components, workflows, and agencies<br>Install pre-built packages from the registry. Publish your own.<br>components<br>Components<br>Reusable capability extensions. Install with one command, invoke with component: and typed inputs.<br>scrapersearchbrowserembeddingttsemail<br>$ kdeps registry install...