workman
YAML agents wired into DAG workflows, run from a local CLI against whichever model provider you like.
Tech stack
- Languages
- TypeScript
- Frameworks
- Node.js
- Ink
- Zod
- Drizzle
- Infra
- libsql
- Turborepo
- AI tooling
- Anthropic Claude
- OpenAI
- Gemini
- Ollama
workman is agent orchestration without a cloud dashboard in sight: define an agent in YAML, wire agents into a workflow as a DAG, and run the whole thing locally from a CLI or a terminal UI, against whichever LLM provider suits the job.
What it does#
- Declarative agents. Agents are YAML files with a schema validated by Zod, not code you have to maintain per provider.
- DAG workflows. Multi-step workflows compose agents with explicit dependencies, so a research step can fan out and a synthesis step can wait.
- Batteries included. Ships with 100 agent definitions and 38 workflow templates across 7 categories, enough to be useful before you write a line of YAML yourself.
- Provider-agnostic. Anthropic, OpenAI, Gemini and local Ollama models behind one interface.
Where it landed#
I have archived workman, honestly, and marked it accordingly. The core is mature, around 215 tests, per-package CI with coverage, full lint and typecheck gates, but the cost dashboard and cron scheduling never got built, and the agent tooling landscape moved fast underneath it. It remains the project that taught me the most about workflow engines: scheduling, retries, and what a DAG executor owes its user when step four fails at 2am.
The stack#
A TypeScript monorepo on Node 22 with pnpm and Turborepo, Zod for schemas, libsql with Drizzle for state, and an Ink (React for terminals) TUI.
