GPU-native agent-swarm orchestration for NVIDIA's AI stack (NeMo, NIM, Triton, NIXL). Workers run inside NVIDIA OpenShell sandboxes or tmux sessions with kernel-level isolation. Supports 8 agent CLIs (Claude Code, Codex, Kimi, Cursor, nanobot, aider, OpenCode, Copilot).
python3 -m venv .venv && source .venv/bin/activate
pip install -e ".[dev]"
pytest tests/ -vsrc/nemospawn/core/— state management (atomic JSON), models, auth, audit, profiles, settings, plan, lifecycle, costs, snapshot, watcher, adaptive scheduling, skillsrc/nemospawn/cli/— Typer CLI commands (24 command groups)src/nemospawn/gpu/— GPU discovery, NVLink topology, DCGM healthsrc/nemospawn/nemo/— NeMo artifacts, config injection, NVLink-aware schedulingsrc/nemospawn/nim/— NIM deploy pipeline, Triton benchmarkssrc/nemospawn/openshell/— OpenShell sandbox integration, security policies, coordination promptssrc/nemospawn/messaging/— file/ZeroMQ/NIXL transport, inboxsrc/nemospawn/observability/— Prometheus metrics, Grafana dashboards, kanban, web UI (SSE)src/nemospawn/templates/— TOML team templates, launch enginesrc/nemospawn/federation/— cross-cluster SSH spawn, git-annexsrc/nemospawn/hpo/— Optuna TPE/ASHA HPO, fallback samplersrc/nemospawn/ngc/— NGC model registry operationssrc/nemospawn/runtime/— tmux, git worktree, SLURM
- All state:
~/.nemospawn/as atomic JSON (tmp + os.replace) - Agents spawn in tmux sessions or OpenShell sandboxes
- GPU pinning via
CUDA_VISIBLE_DEVICES - Config priority: env var (NEMOSPAWN_*) > config file > defaults
- Coordination prompt auto-injected for both tmux and sandbox agents
- Tests:
pytest tests/ -v— 201 tests across 34 files
- Python >=3.10, Typer+Rich CLI, dataclasses (not Pydantic)
rufffor linting, 100 char line length- No database — filesystem-native state