Manners maketh man. Tailoring maketh agents.
Build focused agents.
Run them as a team.
Give every agent its own home: a soul prompt, plus its own skills, hooks, and MCP servers. kman keeps them isolated and runs the whole team from one CLI.
# install the CLI$ bun install -g @unliftedq/kman # create a specialized agent$ kman agent create coder # dispatch it on a mission$ kman -a coder run --task "Refactor auth."Works with the agent runtimes you already use. Right now that's Claude Code and Copilot CLI.
kman never talks to a model itself. It sits on top of the runtimes you already know and sends each job to the agent built for it.
A team beats one big assistant.
Pile every skill and tool into a single assistant and it slows down, runs up the bill, and loses the thread. A handful of narrow agents keeps each run focused and cheap, and keeps mistakes from spreading.
Context stays small
Each agent sees only its own ~/.kman/agents/<name>/ directory. Skills never pile into one giant catalog, so a coder never pays tokens for a researcher's tools.
Focused by design
One soul prompt and a short, curated set of tools. A small surface area keeps the model on the work it's actually good at.
Contained blast radius
Permissions, MCP servers, and hooks are scoped to each agent. Your web-browsing researcher never gets the release bot's write access.
Backend-agnostic
The same agent profile runs on Claude Code or Copilot CLI, so you can switch runtimes without rebuilding your agents.
Plain files on disk
Every agent is just data on disk. Version it, diff it, and share it like any other config file.
Built to compose
Pipe one kman run into the next today. The same isolation boundary will power richer multi-agent flows as they land.
From a fresh machine to a working agent.
Scaffold a named agent with its own profile, soul, skills, and hooks.
$ kman agent create coderEdit soul.md to set the system prompt and shape how the agent thinks.
$ $EDITOR ~/.kman/agents/coder/soul.mdRun a one-off task, or let your agents find and call each other over MCP.
$ kman -a coder run --task "Ship it."Stop building one giant agent that has to know everything. Build a team of small agents that each do one thing well, and let kman put the right one on the job.
Read the docs.
Getting Started
Prerequisites, install, and your first agent.
Concepts
Agents, souls, profiles, backends, isolation.
Architecture
How a run is assembled and launched.
CLI Reference
Every command, flag, and exit code.
Multi-Agent Dispatch
Let agents discover and call each other.
Configuration
Global defaults and environment variables.
Dispatch your first agent.
Install the CLI, create an agent, and put a specialist to work in three commands.
# via bun$ bun install -g @unliftedq/kman # or via npm$ npm install -g @unliftedq/kman $ kman doctor