Usage

Inclavate runs three processes: the orchestrator (routes inference), one or more nodes (each serves a range of transformer layers), and the dashboard UI.

Once installed, dppan, dppan-orchestrator, and dppan-node are on your PATH — run them from any directory. (If you're running from an un-installed bundle instead, prefix the binary with ./ on macOS/Linux or .\ on Windows.)

Single machine

Open three terminals.

1 — Orchestrator

dppan-orchestrator --port 8001 --admin-token mysecret

2 — Node (joins with a model)

dppan join --orchestrator http://localhost:8001 --model llama3.2:1b

3 — Dashboard

dppan ui --orchestrator http://localhost:8001 --port 3000

Open http://localhost:3000.

What a launch looks like

Each process greets you with the Inclavate banner — version, platform, where it's listening, and quick links. Starting dppan-orchestrator --port 8001:

     ◆   ██╗███╗   ██╗ ██████╗██╗      █████╗ ██╗   ██╗ █████╗ ████████╗███████╗
    ╱ ╲  ██║████╗  ██║██╔════╝██║     ██╔══██╗██║   ██║██╔══██╗╚══██╔══╝██╔════╝
   ◆   ◆ ██║██╔██╗ ██║██║     ██║     ███████║██║   ██║███████║   ██║   █████╗
    ╲ ╱  ██║██║╚██╗██║██║     ██║     ██╔══██║╚██╗ ██╔╝██╔══██║   ██║   ██╔══╝
     ◆   ██║██║ ╚████║╚██████╗███████╗██║  ██║ ╚████╔╝ ██║  ██║   ██║   ███████╗
         ╚═╝╚═╝  ╚═══╝ ╚═════╝╚══════╝╚═╝  ╚═╝  ╚═══╝  ╚═╝  ╚═╝   ╚═╝   ╚══════╝
  Run AI privately, across your own devices  ·  Distributed LLM Inference

  orchestrator · routes inference across nodes
  version    0.7.4
  platform   macos-arm64
  system     8 cores
  listen     0.0.0.0:8001
  gateway    0.0.0.0:9001
  mode       local
  database   none

  Docs docs.inclavate.io · Site inclavate.io · Feedback support@inclavate.io · Follow @inclavateai
  © 2026 Inclavate · Free to run · Proprietary
  Legal inclavate.io/privacy · /terms · /acceptable-use · /data-retention

The banner shows only in an interactive terminal — piped output, daemons (systemd/Docker), and CI stay clean. Turn it off anywhere with --no-banner or INCLAVATE_NO_BANNER=1.

Multi-machine

Machine A runs the orchestrator; other machines join as nodes. Nodes need only outbound TCP to ports 8001 and 9001 on Machine A — no inbound ports, no public IP.

Machine A — orchestrator + UI

dppan-orchestrator --port 8001 --admin-token mysecret
dppan ui --orchestrator http://localhost:8001 --port 3000

Machine B, C, … — use Machine A's LAN IP

dppan join --orchestrator http://192.168.1.10:8001 --model llama3.2:1b

Each node auto-detects its GPU, resolves the model via Ollama, and receives a layer range from the orchestrator automatically.

Large models (sharded download)

For models bigger than one machine — say Llama 3.3 70B split across three nodes — each node can download only the layers it will serve with dppan pull --layers, then serve them via dppan join --gguf. Full walkthrough, flag reference, and split-planning tips: Sharded download.

Chat (OpenAI-compatible)

curl -X POST http://localhost:8001/v1/chat/completions \
  -H "Content-Type: application/json" \
  -d '{"model":"llama3.2:1b","messages":[{"role":"user","content":"Hello!"}],"max_tokens":200}'

See the API reference for streaming, sessions, and the node/metrics endpoints.

Local multimodal run

Download the model's projector sidecar once, then attach image/audio files or ordered video frames to standalone local inference:

dppan pull hf:owner/model:Q4_K_M --projector-only
dppan run --model hf:owner/model:Q4_K_M \
  --image photo.jpg \
  --audio question.wav \
  --video-frames frame-0.jpg@0 \
  --video-frames frame-1.jpg@1000 \
  --prompt "Describe these inputs"

--image, --audio, and --video-frames are repeatable. Projectors are discovered from their exact pull metadata; use --projector /path/to/mmproj.gguf for a custom location. In interactive mode, command-line attachments belong to the first turn and remain in later conversation history.

The dashboard paperclip exposes only modalities supported by the selected model. Attachment binaries are SHA-256-addressed in browser IndexedDB while localStorage contains metadata only. This lets saved conversations reconstruct historical media after a dashboard reload or orchestrator KV-cache reset.

Common flags

All three binaries also accept --no-banner (or INCLAVATE_NO_BANNER=1) to suppress the startup banner.

dppan-orchestrator

Flag	Env	Default	Description
--port	PORT	required	REST API port
--admin-token	DPPAN_ADMIN_TOKEN	none	Enables the Admin tab in the dashboard
--models-config	DPPAN_MODELS_CONFIG	config/models.toml (cwd, else next to the binary)	Model registry path
--log-level	LOG_LEVEL	info	trace / debug / info / warn / error

dppan join

Flag	Default	Description
--orchestrator	required	Orchestrator URL, e.g. http://localhost:8001
--model	required	Ollama model name or path to a .gguf file
--gguf	none	Pre-downloaded GGUF (full file or dppan pull shard) — skips Ollama entirely
--layers	auto	Layer range to serve, inclusive both ends like pull (e.g. 0-15 = 16 layers; shards default to their own range)
--log-level	info	Verbosity

dppan pull

Flag	Default	Description
<MODEL>	required	llama3.3:70b (Ollama) · hf:owner/repo[:QUANT] · inclavate:slug
--layers	all	Layer range to download, inclusive (e.g. 0-26)
--orch	off	Include orchestrator tensors and auto-download a published projector
--projector-only	off	Download only the orchestrator projector sidecar
--projector <FILE>	auto	Select an exact projector instead of automatic F16/BF16 selection
--projector-out <PATH>	model store	Override the projector output path
--out	~/.dppan/models/…	Output shard path
--verify	off	Check every tensor's sha256 against the integrity manifest
--remote	off	Range-fetch even if the full blob exists locally

Also: dppan models lists the curated catalog with local availability, and dppan manifest <gguf> generates a per-tensor integrity manifest from a full model file. Projectors reside only on the orchestrator under ~/.dppan/models/projectors/<canonical-model-id>/; worker pulls remain text-only. An adjacent .dppan.json record lets the orchestrator associate the sidecar with catalog and local model records by exact source identity. For a custom location selected with --projector-out, set projector_path in the model's config/models.toml entry.

dppan ui

Flag	Default	Description
--orchestrator	required	Orchestrator URL
--port	3000	Dashboard port
--host	127.0.0.1	Bind address — use 0.0.0.0 for LAN access

GPU troubleshooting

• "CUDA error: the provided PTX was compiled with an unsupported toolchain" — the NVIDIA driver is too old; update to 576.02 or newer.
• cudart64_*.dll / libcudart.so not found — install CUDA Toolkit 12.x. On Linux, add /usr/local/cuda/lib64 to LD_LIBRARY_PATH.
• Falls back to CPU instead of GPU — run nvidia-smi; the GPU must be Turing or newer (GTX 10xx and older are unsupported).
• macOS GPU unused — use the macos-arm64 build on Apple Silicon (Intel Macs run CPU-only).