Getting started

This page covers installing DRAGON, launching it, and running the first sessions. DRAGON is a desktop application: a Tauri v2 shell that supervises the dragond daemon as a bundled sidecar and renders the frontend in the OS-native webview.

Install

Download the installer for your platform from the release artifacts and run it. Per-OS bundles are produced for each release:

The installer bundles the daemon. There is no separate service to install and no terminal window — the shell spawns and supervises dragond automatically. Installers are approximately 8 MB for the NSIS setup, against roughly 150 MB for an equivalent Electron application.

Air-gapped and SCIF installs use these same signed installers, moved by sneakernet. A disconnected install never contacts the update endpoint. See licensing and distribution.

Connected updates

Connected installs self-update over a signed channel. The updater downloads the new bundle, verifies its detached signature against the public key baked into the application, then applies it. Disconnected installs upgrade by installing a newer signed installer by hand.

Launch

Start DRAGON from your application launcher. On start the daemon resolves its data directory, creates the logs/, audit/, and rag/ subdirectories, loads the embedded device profiles and prompt templates, opens its SQLite store and audit log, and binds the loopback WebSocket on 127.0.0.1:7717. The frontend connects to ws://127.0.0.1:7717/ws.

The daemon binds loopback only and is not routable by design. Its Host and Origin headers are hardened against DNS-rebinding and cross-origin connections.

Open an SSH session

Create a session against an SSH host with a device profile. Pick the profile that matches the target — cisco-ios, cisco-nxos, juniper-junos, linux-generic, or generic-fallback — so structured capture segments output correctly.

The first connection to an unknown host raises a TOFU host-key prompt carrying the OpenSSH SHA-256 fingerprint. Accept to append the key to known_hosts; the daemon trusts it on every later connection. A known host presenting a changed key hard-fails as a possible MITM and is not promptable.

In v0.1.0 the daemon wires the SSH password authentication path from the wire protocol. Public-key, agent, and keyboard-interactive auth are supported by the broker but not yet plumbed end to end.

Open a serial session

Enumerate serial ports to populate the picker. DRAGON lists ports with USB friendly names (FTDI, Prolific, CH340). Open a port with the baud, parity, data-bit, stop-bit, and flow settings the console expects. The broker supports the break signal for console recovery procedures.

Save a session

Save transport parameters, the device profile, and a credential reference into the saved-session tree of folders and entries. Saved passwords are written to the OS keychain; the store keeps only an opaque reference, never the secret. Reopen a saved session to dial it without re-entering parameters.

Configure inference

DRAGON runs inference in one of two modes:

• Embedded — a bundled llama.cpp llama-server child process over loopback. This is the offline default and is selected at daemon launch with -llama-bin and -llama-model pointing at a GGUF model file on disk or USB.
• Endpoint — any OpenAI-compatible server: LM Studio, Ollama, vLLM, llama-server, or a cloud provider. Set the base URL, model name, embedding model, and API key in settings.

Endpoint mode hot-swaps live from settings — the client, embedder, and redaction strictness reconfigure without a restart. Switching to embedded mode requires a daemon restart with the -llama-bin and -llama-model flags.

When the inference endpoint is non-loopback, DRAGON automatically applies strict redaction, which adds a high-entropy token heuristic on top of the structural rules.

Ask the copilot

With a session open and inference configured, ask the copilot a question about the current session. The orchestrator assembles a context window — recent structured events, the device profile, current mode, and RAG retrieval results — redacts every piece, calls the model, and returns an answer. Malformed model output degrades gracefully to plain text and never produces a staged suggestion.

Staged next-command suggestions land in the staging strip above the input line, each tagged read-only, config-impacting, or destructive. Accept a suggestion to place it in the input line. You send it. DRAGON never transmits it for you.

Add a RAG corpus

Register a documents or configs source by pointing DRAGON at a directory. The ingestor walks the path (Markdown, TXT, HTML, and PDF for docs), redacts the text, chunks it structurally, embeds it, and stores the vectors. Retrieval is hybrid — vector search fused with BM25 keyword search — and insights cite their sources.

Session history is ingested automatically from completed command records, post-redaction. RAG ingestion is confined to allowed roots; the corpus cannot be pointed at arbitrary host files.

Review the audit log

Tail the audit log from the audit view. Every suggestion, acceptance, dismissal, redaction event, and model call is recorded in an append-only, hash-chained JSONL log under audit/. Export it as JSONL or CSV for compliance ingestion. The format is a public specification — see redaction and audit.