Home

Deployment Modes

Bodhi App ships in two flavors. Both run the same server, expose the same APIs, and use the same OAuth2 authentication — but they target very different workflows. This page explains the differences so you can pick the right one.

At a glance

Tauri Desktop Docker (single-tenant)
Who runs the OS process? You — it's a desktop app on your machine. A server you control (your VPS, a homelab box, a cloud VM, your team's NAS).
Where does data live? ~/.bodhi/ on your home directory. A Docker volume you mount (typically /data inside the container).
Who can reach the UI / API? Just you, on localhost:1135. Anyone you allow — bind to a port, put a reverse proxy in front for TLS.
Single user or many? Single user (you). Multiple users with separate accounts and roles.
OAuth required? Yes — first launch walks you through admin sign-in. Yes — first user becomes admin, others request access.
Updates? Download a new installer (Bodhi notifies you in-app). docker pull the new tag and recreate the container.
Best for... Personal use, dev workstations, "I just want to run a model on my laptop." Team usage, always-on inference, GPU servers, sharing local models with co-workers.
GPU acceleration Yes — Metal on macOS, CUDA on Windows/Linux (variant-dependent). Yes — pick a Docker variant (CUDA, ROCm, Vulkan, MUSA, Intel, CANN).

Tauri Desktop

Bodhi runs as a native desktop app on macOS (Intel + Apple Silicon), Windows, and Linux. Under the hood it's the same server binary — wrapped in Tauri, with a system-tray icon and an embedded UI shell.

What's different from Docker:

  • The server starts when you launch the app and stops when you quit. There's a tray icon to keep it running in the background.
  • Data lives in ~/.bodhi/ (logs, database, downloaded model aliases). GGUF files themselves go in your HuggingFace cache (~/.cache/huggingface/hub/) so they're shared with other tools.
  • The app pops a browser to http://localhost:1135 automatically.
  • Only you, on this machine, can reach it. No port is exposed externally.
  • macOS: encryption keys are stored in Keychain. Windows: in DPAPI. Linux: in the secret service or a fallback file.

Pick this if: you want zero ops, a single-user experience, and the lowest possible friction. It's also the fastest way to evaluate Bodhi.

For installation, see Install → macOS / Windows / Linux. The full deployment page lands at /docs/deployment/desktop in a later phase.

Docker (single-tenant)

You run Bodhi as a container — typically ghcr.io/bodhisearch/bodhiapp:latest-cpu or one of the GPU variants — on a server you control. One Bodhi instance serves multiple human users, each with their own login, role, and API tokens.

What's different from Desktop:

  • You manage the process: docker run, docker compose, systemd, Kubernetes — your call.
  • You bind a port (default 1135) and decide who can reach it. Production deployments typically put a reverse proxy (Caddy, nginx, Traefik) in front for TLS termination.
  • Data lives in a mounted volume. You set BODHI_HOME on the container and back up that path.
  • Multiple users can sign up via OAuth. The first one becomes admin; others submit access requests that admins/managers approve.
  • You pick a variant matching your hardware: AMD64 / ARM64 CPU, NVIDIA CUDA, AMD ROCm, cross-vendor Vulkan, Moore Threads MUSA, Intel GPU, or Huawei CANN.

Pick this if: you want a shared, always-on inference server for a team, or you want to run on a GPU box that isn't your daily-driver laptop.

For deployment instructions, see Docker Deployment. The decision-flow page (Tauri vs Docker, with reverse-proxy guidance) lands at /docs/deployment/overview in a later phase.

What both modes share

  • Same APIs. Every endpoint described in API Compatibility works identically on both.
  • Same auth model. OAuth2 PKCE login, the four roles, scoped API tokens — see Auth and Roles.
  • Same MCP infrastructure. Each user has their own MCP instances; admins maintain a shared catalog of pre-registered servers — see MCP Overview.
  • Same model concepts. Local model aliases, GGUF files, API model aliases — see Models, Aliases, and Files.
  • Same swagger. The running app exposes its full schema at /swagger-ui regardless of mode.

Common questions

Can I switch later? Yes — there's no lock-in on either side. Both modes use the same SQLite/Postgres schema (Desktop is SQLite; Docker can be either). The simplest migration is "export aliases and API model configs, re-create on the other side." Direct DB-file portability between Desktop and Docker is not officially supported.

Can I run both at the same time? Yes — they're just two server processes. They won't share data, though.

What about reverse proxies, TLS, rate limits? That's a Docker-mode topic. The full reverse-proxy guide lands at /docs/deployment/reverse-proxy in a later phase. Short version: terminate TLS upstream, forward Host and X-Forwarded-* headers, and set BODHI_PUBLIC_HOST / BODHI_PUBLIC_SCHEME so OAuth callback URLs come back to the right place.

Ready for the next building block? Head to Models, Aliases, and Files to disambiguate the three things "model" can mean in Bodhi.