Orchestrating teams of specialist agents in one workspace.
Crew44 turns the AI agents you already run into coordinated teams. Specialists sit in one workspace, hand off the baton, and accumulate skills with every run. Everything stays on your machine.
Parallel work,
shared context.
The default crew comes with a Partner, an Engineer, a Product Lead, and a Designer. Each owns a role, a model, and its own skills folder. They debate decisions, hand off the baton with a one-line brief, and ship in parallel — instead of one agent re-explaining context to itself all day.
src/export.ts …Long-term memory,
auto-optimized.
Skills and per-project memory live as plain files on disk and persist across every run. A background Partner mines your history on a schedule and proposes new memories, skills, and routing tweaks — always with evidence from real tasks. You click Accept, Edit, Snooze, or Dismiss. Nothing lands without your click. Six months in, the crew knows your codebase better than any single agent ever could.
The best model
for every move.
Every model has tradeoffs — Opus is deep but slow, GPT-5.5 is fast on codegen, a local model is free for review. Crew44 binds each agent to the model that wins at its job, and agents can switch models mid-task: a thinking model for the hard turn, a fast model for the obvious next step. The same skills folder runs across providers, so you’re never locked in.
Reach the crew
from your phone.
A companion mobile client pairs with your daemon over an end-to-end encrypted tunnel. Approve a handover from the couch, nudge an agent from the train, watch a long run land at lunch. The crew keeps shipping while you’re away from the laptop.
Local-first, by design.
Your code, your skills, and your run history never leave your laptop. The daemon binds to 127.0.0.1 and the only outbound traffic is whatever your underlying coding agent already makes. State is plain files at ~/.crew44/ — grep it, version it, delete it. Auth is a per-launch bearer token; /health is the only unauthenticated endpoint. If you don’t like it, read the daemon — it’s one Go process.