swarmr

Teams of Deep Agents domain specialists, exposed two ways: a streaming terminal CLI and an MCP server.

This is the core distribution. It ships no team. Teams live in their own repositories, depend on this one, and are found through installed metadata — pip install and pip uninstall are the whole lifecycle.

Install

Core and the team must land in the same environment; discovery reads installed distribution metadata, not paths.

uv tool install "swarmr[all]" --with-executables-from swarmr-kube

all pulls in every first-party team; swarmr[kube] or swarmr[blame] picks one. --with-executables-from is easy to miss: without it only core's teams and teams-mcp reach your PATH. Name each team package that ships commands (swarmr-kube ships incident-credentials; swarmr-blame ships none).

From a checkout instead of PyPI:

uv tool install ../swarmr --with ../swarmr-kube --with ../swarmr-blame \
  --with-executables-from swarmr-kube

Then a model key, in the directory you run from:

printf 'KIMI_API_KEY=sk-…\n' > .env && chmod 600 .env

Loading walks up from the current directory, then falls back to ~/.env; KIMI_ENV_FILE overrides both.

CLI

teams --list
teams --target <team>
teams <team> "the symptom, as prose"

Output shows delegation live: the plan, parallel dispatches, each tool call with a digest of its result, every subagent's verdict attributed by name, per-agent token counts.

MCP

{
  "mcpServers": {
    "swarmr": {
      "type": "stdio",
      "command": "teams-mcp",
      "cwd": "/abs/path/to/your/working/dir",
      "timeout": 0
    }
  }
}

cwd is where the server looks for .env. Tools: start_<team> for every installed team, plus check_task and list_tasks. Runs take minutes, so start_* returns a job id immediately and check_task long-polls — returning the moment a specialist is dispatched or reports, with only the trail steps you have not seen.

Install a team, restart the server, and its tool appears. No configuration changes.

Design notes and internals: CLAUDE.md.