Offline-first AI agent orchestration. Hand off a requirement, walk away, come back to merged code.
NXD decomposes a natural-language requirement into stories, dispatches them in parallel waves to local-LLM-powered agents, runs code review + QA, and merges — fully offline, no API keys, no rate limits.
- No API costs — run as many agents as your hardware allows
- Full privacy — code never leaves your machine
- Works anywhere — planes, off-grid, behind firewalls
You need: Go 1.26+, Ollama, tmux, Python 3 for MemPalace.
# 1. Start Ollama (skip if already running as a service)
ollama serve & # or: brew services start ollama
# 2. Pull models
ollama pull qwen2.5-coder:14b # planner + reviewer (~9 GB)
ollama pull gemma4:e4b # coder (~6 GB)
# 3. Install NXD + MemPalace
go install github.com/tzone85/nexus-dispatch/cmd/nxd@latest
pip install -r requirements.txt # MemPalace
# 4. Run inside a git repo
nxd init
nxd req --background "Build a REST API for user management"
nxd req-logs <req-id> # tail the daemon log
nxd dashboard # TUI; add --web for the browser dashboardThe first run takes a few minutes while Ollama warms the models. nxd req --background self-daemonizes the pipeline; without it, nxd req plans then exits and you must run nxd resume <req-id> to dispatch agents. Follow progress via nxd req-logs <req-id> or the dashboard.
| Platform | Build | Read-only commands | Full agent pipeline |
|---|---|---|---|
| macOS (Apple Silicon, Intel) | native | yes | yes |
| Linux (x86_64, arm64) | native | yes | yes |
| Windows 10/11 (native, no WSL) | GOOS=windows go build |
yes | no — requires tmux |
| Windows + WSL2 (Ubuntu/Debian) | inside WSL | yes | yes |
NXD's agent execution pipeline depends on tmux for session isolation, recovery,
and live inspection, and tmux has no native Windows port. On native Windows the
binary still builds and runs — nxd status, nxd dashboard, nxd doctor,
nxd config, nxd events, and the report/metrics commands all work — but
nxd req / nxd resume will fail the tmux check with a clear pointer to
WSL2. The recommended Windows install path is WSL2 with Ubuntu, where the
Linux instructions in Quick start apply unchanged.
go install github.com/tzone85/nexus-dispatch/cmd/nxd@latest
# Resulting binary lives at %USERPROFILE%\go\bin\nxd.exe — ensure that
# directory is on PATH (it is by default if you installed Go via the MSI).
nxd init
nxd doctor # Will warn on the tmux check; everything else should pass.State on Windows lives under %USERPROFILE%\.nxd\ by default. The devdb
provider's Docker fallback dials tcp://localhost:2375 (enable "Expose daemon
on tcp://localhost:2375 without TLS" in Docker Desktop) — or set DOCKER_HOST
explicitly. To pick a different host shell for the native runtime's
run_command tool and user metric/migration commands, set NXD_SHELL=pwsh
(or any shell available on PATH); the default is cmd.exe.
# In an elevated PowerShell, one time:
wsl --install -d Ubuntu# Inside the Ubuntu WSL shell:
sudo apt update && sudo apt install -y tmux git build-essential
# Install Go 1.26+ (apt's package may lag — see https://go.dev/dl).
go install github.com/tzone85/nexus-dispatch/cmd/nxd@latest
nxd init && nxd doctor| Command | What it does |
|---|---|
nxd init |
Create ~/.nxd/, generate nxd.yaml, check Ollama |
nxd req "<requirement>" |
Submit a requirement; planning + dispatch |
nxd req --background |
Same, but self-daemonize; tail with nxd req-logs |
nxd status |
Requirements + stories overview |
nxd dashboard [--web] |
Live TUI / browser dashboard |
nxd doctor |
Preflight checks (Go, git, tmux, Ollama, config) |
Full CLI reference: docs/reference/cli-reference.md.
| If you want to... | Read |
|---|---|
| Walk through your first run step-by-step | Getting Started |
| Pick models for your RAM / VRAM | Model Selection |
| Tune Gemma 4 or function calling | Gemma 4 Guide |
Edit nxd.yaml confidently |
Configuration |
| Understand how the pipeline works | Architecture Deep Dive |
| Use ephemeral per-story Postgres DBs | docs/guides/configuration.md#devdb |
| Switch to cloud mode (Anthropic, GitHub PRs) | Configuration → models / merge |
| Fix a problem | Troubleshooting |
| See every event type the system emits | Event Reference |
| Build from source on macOS | docs/guides/getting-started.md#local-build |
cmd/nxd/ CLI entry point (cobra)
internal/agent/ Role definitions, prompts, complexity scoring
internal/cli/ Cobra command implementations
internal/config/ YAML loader + validation
internal/dashboard/ TUI (Bubbletea), web (WebSocket)
internal/engine/ Planner, dispatcher, reviewer, QA, merger, watchdog
internal/git/ Worktrees, branches, local merge, GitHub PRs
internal/llm/ Ollama / Anthropic / OpenAI clients
internal/runtime/ Pluggable runtimes (gemma native, Aider, Claude Code, Codex)
internal/state/ Event store (JSONL) + SQLite projections
internal/tmux/ Session lifecycle
Architecture deep dive: docs/guides/architecture.md.
make build # build nxd
make test # tests with race + coverage
make install # install to $GOPATH/binPatterns and ideas borrowed from Gastown, Beads, Dolt, Hungry Ghost Hive, and Wasteland. Worth a read if you're into agent orchestration.
Made with ❤️ by Vortex Dispatch (NXD is the offline-first sibling). Remove this line freely.