A decentralized CI/CD platform that turns idle machines into a distributed build fleet.
Traditional CI/CD runners are expensive and centralized — you rent someone else’s machine while your own hardware sits idle. Hive inverts that model: anyone can install a lightweight agent daemon on a spare machine and contribute it to a shared compute grid.
A central FastAPI coordinator schedules work across contributed machines using weighted fair queuing, streams container logs back in real time over WebSockets, and meters CPU and RAM per second against a credit economy. Contributors earn credits by running jobs; requesters spend them.
Owns all state and every decision. A 20-endpoint REST API behind JWT auth, an async scheduling loop on a 5-second tick, per-second metering against balances, and a WebSocket hub fanning updates out to dashboards. It never executes user code itself.
FastAPI · SQLAlchemy 2.0 · PostgreSQL
A cross-platform daemon that registers, heartbeats CPU and RAM telemetry, and polls for work matching its free capacity. It claims a job, downloads the payload, runs it in an isolated container and streams stdout upstream in chunks as it goes.
Python · Docker SDK · psutil · PyInstaller
Job submission, a live grid of every node’s status and utilization, per-step pipeline inspection, streaming log panels, credit history and analytics — all driven by the socket rather than a refresh button.
Next.js 15 · React 19 · Tailwind v4
A channel-based WebSocket manager pushes grid-wide snapshots to /ws/dashboard. Queue depth, running jobs, agents online and credit flow update themselves — nothing polls.
Upload a zip or point at a Git repo, pick a simple command or a multi-step pipeline, and set the cores, RAM and array size. The cost preview breaks out the base fee against metered compute before anything is dispatched.
The agent ships as a system-tray desktop app so a non-technical contributor can bring capacity online without a terminal. Cores and RAM are capped per node, and logs stream upstream while the container is still running.
Balance influences priority, but wait time compounds continuously — so a long-queued job from a low-balance user eventually outranks a wealthy one. Agents sort independently by remaining capacity, descending: a best-fit placement.
weight = wait_seconds × (1 + balance × 0.05)Miss the heartbeat and the agent is marked offline; anything running on it is detected as orphaned and requeued with its binding cleared. A node can be unplugged mid-build without losing work.
system_error and abandoned stay separate terminal states, so an agent crash never looks like an unclaimed job.
Every movement writes a transactions row instead of mutating a counter, so any balance reconstructs from history.
Each step carries its own status, timestamps and captured output, so a failure names the stage rather than dumping one undifferentiated log.
array_size > 1 fans out into N child jobs sharing a single payload.