JellyClaw structures agents the way companies structure people: one accountable chain from goal to task, defined entirely in YAML.
Work flows down the hierarchy; results and accountability flow back up.
Owns the goal. It plans: breaking your objective into department-sized pieces, sequencing them, and deciding what matters tonight. It delegates each piece to the right department head, then assembles their results into the report that lands in your Telegram.
Own an area — dev, research, ops. A head routes incoming work to the worker best suited for it, reviews what comes back before accepting it, and sends anything below the bar back down with notes. Nothing reaches the CEO unreviewed.
Own execution. A worker does exactly one kind of task — write code, review a diff, summarize sources — using the model and tools its YAML gives it, and reports the result to its head. Small scope, sharp focus.
Model, role, tools — that's an agent. No code, no plugins, no framework to learn.
# planned format for the CLI MVP
agents:
code-reviewer:
role: worker
model: qwen2.5-coder:14b # any Ollama model, or claude for hard tasks
tools: [git, shell]
research-writer:
role: worker
model: llama3.1:8b
tools: [web, files]Two ready-made offices ship with the CLI MVP. Start from one, then reshape it.
A coding department: workers draft changes and review each other's diffs overnight; the head signs off before anything reaches you.
team: dev-team
departments:
engineering:
head: llama3.1:8b
workers: 2
report:
channel: telegramA research-and-outreach pair: one department digs into prospects and competitors, the other drafts the follow-ups for your morning review.
team: sales-office
departments:
research:
head: llama3.1:8b
workers: 2
outreach:
head: llama3.1:8b
workers: 1Watch the repo to catch the CLI MVP the moment it lands.