Single-node scheduling for shared machines
Give one Linux machine a real job scheduler
A small CLI and a local daemon to submit, schedule, and inspect GPU or CPU jobs.
- Single-node by design
- GPU-aware scheduling
- tmux-backed execution
- MCP-ready for agents
Why it exists
When one machine is no longer just yours
Ad-hoc tmux and GPU etiquette fall apart once the box is shared.
Without discipline
- Long jobs collide with interactive work.
- Failures scatter across logs with no state to recover.
- GPU rules live in tribal knowledge, not policy.
With gflow
- One daemon owns state and the queue.
- Every job is inspectable and recoverable.
- Resources and dependencies are declared up front.
How it works
Four commands, one controlled run
gflowd upStart the daemongbatch --gpus 1Submit a jobgqueueWatch it schedulegjob log <id>Follow the runCapabilities
Built for daily workstation use
Queueing & lifecycle
Submit, hold, cancel, update, and redo jobs.
GPU-aware scheduling
Request GPUs, share them, and cap VRAM.
Workflow composition
Chain dependencies, arrays, and sweeps.
Operational visibility
Read state as table, tree, JSON, CSV, or YAML.
Recoverable execution
Every job runs in its own tmux session.
Automation & AI
Drive the scheduler through a local MCP server.
Where it fits
Where it fits
Shared lab GPU server
Coordinate many researchers on one box.
Solo research machine
Keep long experiments structured and restartable.
Local eval pipelines
Chain prep, train, benchmark, and report.
Documentation
Start where you are
AI integration
Turn scheduler operations into agent tools
Run gflow as a local stdio MCP server; agent CLIs can inspect queues and drive workflows.
gflow mcp serveStart scheduling the machine you already have
The docs double as an operator's handbook.