apra-fleet

apra-fleet is an open-source tool that lets you orchestrate a fleet of AI agents across the machines you already own. Think of it as Kubernetes for AI agents: it handles scheduling, credentials, isolation, and observability so you can run multiple agents on different devices and LLM providers from a single control plane. You install it via npm, bring your own API keys, and start registering machines as "members" — each member pairs a real device (laptop, GPU box, VM) with a provider CLI (Claude, Gemini, Codex, Copilot, Antigravity, or local models).

What makes apra-fleet stand out is its layered architecture. It's an MCP server plus CLI, so any MCP-capable agent (like Claude Code or Gemini) becomes the orchestrator. You drive it in plain language: "Register a local member called doer. Register another called reviewer. Pair them." Workflows are durable and resumable — they have execution journals, turn budgets, stall detection, and crash watchdogs. A key innovation is the cost model: you can explore a workflow with an LLM orchestrating (token-hungry but flexible), then harden it into a deterministic program where shell, git, and file steps run at zero tokens, and models are only invoked at judgment nodes. The project's own e2e tests show costs dropping from $0.46–$3.05 per run to roughly $0 after hardening.

This tool is built for developers and teams who need to run multi-agent workflows across heterogeneous hardware without vendor lock-in. It's especially useful for autonomous software development — the flagship workflow "fleet-sprint" plans, develops, reviews, deploys, and tests code in cycles, and the project itself is built by the product. If you're tired of single-agent demos and need a real operations layer for an agentic workforce, apra-fleet is worth a look.