collar

DAG-first AI agent harness — Entity→Target:verb(card). Reads .dag files for graph-backed context.

.dog = human spec format · .dag = agent runtime format · .md = last resort

Collar is forked from NousResearch/hermes-agent, rebuilt as a DAG-first harness. It reads .dag files as its behavioral ground truth. Skills auto-compile from .dog specs into .dag edges. No prose config, no baked-in prompts — just DAG paths.

Created by Justin DiClemente (@logohere).

Install

git clone https://github.com/specdog/collar.git
cd collar
python3 -m venv .venv && source .venv/bin/activate
pip install -e .
npm install -g dotdog

Requires Python >= 3.11.

Quick Start

collar              # interactive chat
collar chat -q "..."  # single query
collar setup          # first-time setup wizard

No API key? Use OAuth

collar auth add openai-codex   # ChatGPT Plus/Pro
collar auth add xai-oauth      # xAI SuperGrok
collar auth add qwen-oauth     # Qwen Plus/Pro
collar auth add nous           # Nous Portal

Then collar model to pick your provider. Or set OPENROUTER_API_KEY in ~/.dag/.env for pay-per-token access to 200+ models.

Architecture

Layer	Source	Format	Loader
System prompt	`dags/*.dag`	DAG-path plain text	`_load_dag_text()`
Skills	`skills/*/SKILL.dag`	DAG-path plain text	`skill_view()`
Projects	`*.dag`	JSON v2/v3	MCP dotdog

Evaluate the workflow

Compare complete runs, task correctness, model/provider settings and repeated measurements before making token or timing claims. Compression of one artifact is not proof of lower total cost or better agent outcomes.

Read the Collar product notes for its relationship to Dotdog and the evidence boundaries.

Links

GitHub · dotdog · specdog org