apra-fleet

What is apra-fleet?

apra-fleet is an open-source MCP-compatible control plane for running many AI coding agents across machines you already own. Its own framing is that it does for agents what Kubernetes did for containers: scheduling, credentials, isolation and observability for a fleet whose members can run locally or over SSH and can mix every LLM provider at once. Members are registered in plain language, paired into roles such as doer and reviewer, and driven through single autonomous multi-hour fleet-sprint runs with stall detection and crash recovery. It installs from npm as `@apralabs/apra-fleet`, is Apache-2.0, and has no paid tier — you bring your own LLM keys.

Key features

• Fleet of agent 'members' registered in plain language and paired into roles (for example a doer and a reviewer), with a fleet manifest and live console
• Members run locally or over SSH across Windows, macOS and Linux, mixing multiple LLM providers in one fleet
• Autonomous fleet-sprint runs that go multi-hour and unattended, with stall detection and crash recovery
• Tier-based cost routing and cross-provider review, so cheaper models can carry work that does not need a premium one
• Out-of-band credential handling: secrets are typed into a separate terminal, encrypted at rest and resolved server-side at execution time, so no model ever sees them
• Credentials scoped per member with TTL expiry and per-credential network egress policy, plus composed provider-native allow-listed permissions
• MCP-compatible, installed with `npm install -g @apralabs/apra-fleet` (Node.js 22+) or standalone installer binaries for each platform
• Apache-2.0 with no paid tier; the project states it builds itself using its own autonomous fleet-sprint runs

Who it's for

• Putting idle machines — a spare desktop, an old laptop, a box reachable over SSH — to work as agent workers instead of running everything on one laptop
• Running long unattended coding jobs that outlive a single terminal session, with recovery when a member stalls or crashes
• Having one agent write and a second agent on a different provider review the work, as a cross-provider check
• Cutting spend on a multi-agent workload by routing tasks to cheaper model tiers rather than sending everything to a premium model
• Giving agents access to credentialed systems without putting the secrets in a model's context

When not to use it

Skip it if you want something managed or finished. It is v0.4.0, parts of its own feature list are labelled preview or early, there is no hosted tier or support contract, and you take on operating the machines yourself — Node.js 22+ on each host, SSH reachability, and every provider key and subscription supplied and paid for by you. If your agent work fits comfortably on one machine, the fleet layer is overhead you do not need yet.

FAQ