kdeps agent

An autonomous LLM REPL you run locally. Tool use, persistent memory, multi-step reasoning, skills, goals, and a judge panel - against a local model, no API key, fully offline. Whole workflows and components register as callable tools; the model decides what to invoke.

kdeps            # bare REPL - built-in tools only
kdeps ./my-agent # also load that agent's workflows/components as tools

Not this? If you want a deterministic request→response pipeline (same input, same execution path, safe to run unattended), that's kdeps workflow. To orchestrate several agents as one system, kdeps agencies.

---

Running kdeps with no arguments starts a bare REPL with no workflow tools - built-in tools (web_search, bash_exec, file ops, memory, etc.) are still available. Pass a path to also load workflows and agencies as tools.

Starting the agent loop

kdeps                              # runs the agent loop REPL
kdeps --model llama3.2 --system "You are a DevOps assistant."  # override model/system prompt
kdeps --skill ~/.kdeps/skills/     # load skill files
kdeps --resume <session-id>        # continue a saved session

Single workflow vs folder

kdeps ./my-agent/     # registers the workflow as an LLM-callable tool (named after metadata.name)
kdeps ./agents/       # registers every workflow and agency in the folder as a separate tool

When you point at a folder, kdeps discovers every workflow and agency file inside it (recursively). Each becomes a separate tool. The tool name is metadata.name from the workflow's manifest - not the filename.

Target	Workflow/agency/component tools registered
No path	None (built-in tools only)
Single workflow file/dir	One tool (metadata.name) + one per component
Single agency file	One tool (agency metadata.name)
Folder	One tool per workflow/agency found recursively + component tools

How it works

Given a workflow whose metadata.name is my-agent, running kdeps ./my-agent/ gives the LLM one tool named my-agent. When it calls that tool, kdeps runs the full workflow DAG - every resource in dependency order - and returns apiResponse.response to the LLM. The LLM can then call more tools or produce a final answer.

Command and flags

kdeps [path] [flags]

[path] is optional. When provided it must be a workflow or agency file or directory.

Flag	Default	Description
--model	KDEPS_AGENT_MODEL or llama3.2:1b	LLM model name
--backend	KDEPS_AGENT_BACKEND or file	LLM backend (file, gguf, ollama, openai, ...)
--base-url	KDEPS_AGENT_BASE_URL	LLM API base URL
--system	(none)	System prompt injected at conversation start
--skill	(none)	Path to a skill file or directory (repeatable)
--prompt	(none)	Path to a prompt templates directory (repeatable)
--resume	(none)	Session ID to resume a previous conversation
--stealth	false	Muted UI - dark gray, model name barely visible (for use in public)
--debug	false	Enable debug logging

# Environment variables (flags override these)
KDEPS_AGENT_MODEL=llama3.2:1b
KDEPS_AGENT_BACKEND=file              # default: local llamafile
KDEPS_AGENT_BASE_URL=http://localhost:11434
KDEPS_STEALTH=1                       # same as --stealth (1, true, or yes)

Examples

kdeps                                            # bare REPL, built-in tools only
kdeps ./my-agent/                                # register one workflow as a tool
kdeps ./agents/                                  # register every workflow in the folder
kdeps ./agents/ --model mistral --system "You are a data analyst."
kdeps --backend gguf --model qwen3.5-4b          # local GGUF model
KDEPS_AGENT_BACKEND=openai kdeps ./agents/ --model gpt-4o
kdeps --resume abc123def456                      # resume a session

More on the agent loop

Topic	Page
Slash commands and auto-detected shell/file mentions	REPL slash commands
Built-in tool catalog, permission modes, lean mode	Built-in tools
Shell execution (!cmd, Ctrl+C / Ctrl+Z, jobs)	Shell execution
Live status line and stall detection during a tool run	Tool execution monitoring
Task decomposition and forward-only goal enforcement	Goal-directed execution
Independent review of the final answer	Judge panel
/model, /context, auto-routing, running local servers	Local model management
Task / team / cron registries (task_*, team_*, cron_*)	Agent registries
One-time permission exceptions for a denied tool call	Approval tokens
Optional turo token reducer	Prompt reduction (turo)
Skills, prompt templates, KDEPS.md instructions	Skills and prompt templates
Pasting, rendering, stealth, sessions, notifications, updates	Agent loop REPL features

Differences from workflow mode

	Workflow mode (kdeps run)	Agent mode (kdeps [path])
Execution	DAG, deterministic	LLM loop, tool-driven
Entry point	metadata.targetActionId	User prompt
Unit of work	Individual resources	Whole workflows
Tools	Functions in chat.tools	One per workflow + one per component + built-ins
Input	One workflow path	Optional file or folder
Conversation	Single run	Multi-turn, persistent JSONL

See also

• Agent loop REPL features - pasting, rendering, sessions, updates
• Skills and prompt templates - context files that teach the agent
• REPL slash commands - full command reference
• Workflow mode - deterministic DAG pipelines
• LLM provider reference - backend config and model names
• Agencies - multi-agent orchestration