kdeps turns one workflow.yaml into an offline API, a binary, an image, or an ISO. You review that file as a pull request. kdeps test runs its evals.
- Workflows - a YAML pipeline is the one way to run kdeps. Feed it from an HTTP API, a bot, a file, or an interactive prompt.
- Agents - a
chatwithloop:is an AI agent with tools and memory, inside the same workflow.kdeps new helper --template assistantstarts one. - Deploy on your infrastructure - ship the same YAML as a Docker image, Kubernetes manifests, a bootable ISO, or one binary.
brew install kdeps/tap/kdeps # macOS or Linux
curl -LsSf https://raw.githubusercontent.com/kdeps/kdeps/main/install.sh | sh # or the install scriptirm https://raw.githubusercontent.com/kdeps/kdeps/main/install.ps1 | iex # Windows PowerShellFor other options, see the installation guide. If you upgrade from v2, see what changed.
Save as workflow.yaml:
apiVersion: kdeps.io/v1
kind: Workflow
metadata:
name: my-agent
version: "1.0.0"
targetActionId: response # the resource whose output is returned
deploy:
apiServer:
portNum: 16395
routes:
- path: /api/v1/chat
methods: [POST]
resources:
- actionId: llm
name: Ask the AI
validations:
check:
- get('q') != '' # reject the request before any model call
error:
code: 400
message: "'q' is required"
chat:
model:
name: llama3.2:1b # local model, no API key
timeout: 60s
context:
prompt: "{{ get('q') }}"
- actionId: response
name: Return the answer
flow:
requires: [llm] # runs after llm
apiResponse:
success: true
response:
answer: get('llm').message.contentRun it and ask a question:
export KDEPS_API_AUTH_TOKEN=dev-token # local testing only
kdeps run workflow.yaml
curl -s -X POST http://localhost:16395/api/v1/chat \
-H "Authorization: Bearer $KDEPS_API_AUTH_TOKEN" \
-H "Content-Type: application/json" \
-d '{"q": "Explain entropy in one sentence."}'Resources run in dependency order. Independent resources run in parallel. The model can write a different answer each time. But for the same request, kdeps always runs the same steps in the same order.
A chat puts its keys in these groups: model, context, harness, tools, rag, and guardrails. Guardrail judges review and revise the answer. tools.list names built-in tools (web_search, calculator, subagent, ...), or auto for all of them. Judges are opt-in.
kdeps test . runs the evals: of a workflow in the process. It runs assertions, judges, and repeated runs against a pass-rate threshold.
For more data, see the workflow docs, examples, and agencies (many agents as one system).
An agent is a workflow whose chat has a loop: block: the model calls tools over many rounds and keeps one conversation per session.
kdeps new helper --template assistant # interactive prompt + chat with loop:
cd helper && kdeps # bare kdeps runs ./workflow.yaml
kdeps --agent-loop # the agent REPL, with slash commands and a model picker# helper/resources/assistant.yaml
chat:
context:
prompt: "{{ input('message') }}"
loop:
tools: [read_file, edit_file, bash_exec] # the tools the agent may call
session: "{{ get('session') }}" # one conversation across promptsFile tools stay inside the working directory. In the REPL, a path outside it asks you first; ~/.kdeps and /tmp are exceptions.
For more data, see the agent docs.
kdeps bundle build . # container image
kdeps export k8s . # Kubernetes manifests
kdeps export iso . # bootable ISO
kdeps bundle package . # my-agent-1.0.0.kdeps
kdeps bundle prepackage my-agent-1.0.0.kdeps --include-models # binary with the modelConnect agents to an LLM server that you host. Then prompts and data stay on your network:
# ~/.kdeps/config.yaml
llm:
backend: openai
base_url: http://192.168.1.50:8000/v1 # your OpenAI-compatible server
models:
- llama3.2kdeps llm wizard builds that server. For more data, see deploy, LLM server, and security.
Read CONTRIBUTING.md. kdeps has the Apache 2.0 license.