@gorlitzer-labs/seldon

 ███████╗███████╗██╗     ██████╗  ██████╗ ███╗   ██╗
 ██╔════╝██╔════╝██║     ██╔══██╗██╔═══██╗████╗  ██║
 ███████╗█████╗  ██║     ██║  ██║██║   ██║██╔██╗ ██║
 ╚════██║██╔══╝  ██║     ██║  ██║██║   ██║██║╚██╗██║
 ███████║███████╗███████╗██████╔╝╚██████╔╝██║ ╚████║
 ╚══════╝╚══════╝╚══════╝╚═════╝  ╚═════╝ ╚═╝  ╚═══╝

The installer for the Seldon stack — a software factory of AI agents. Pick the tools you want from a checklist; each installs via its native method (npm for the Node tools, a shell installer for bifrost, a venv for demerzel).

Install

pnpm add -g @gorlitzer-labs/seldon     # or: bun add -g …  ·  npm i -g …
seldon                 # open the checklist — ↑↓ move · space pick · a all · enter install

Or name the modules directly:

seldon install foundation comb   # install just these
seldon doctor                    # check external deps (tmux, sops, age, tailscale, python)
seldon list                      # everything available

Start the stack

One command brings up the always-on services and prints where to go:

seldon up            # starts the voice (demerzel) + the supervisor (factory watch),
                     # then prints: Voice → http://localhost:8770, how to put agents
                     # to work, rooms, phone access
seldon up --tailnet  # also expose the voice on your tailnet → reach it from your phone
seldon status        # what's running (pids)
seldon down          # stop everything

Voice brain (Qwen or Bonsai): the first seldon up after installing the voice asks once — q for Qwen 3.6-35B (in-process, sharpest, ~20 GB) or b for Bonsai 2-27B (a local llama-server, lighter ~7 GB, tool-capable) — and remembers the choice. Switch anytime with seldon up --brain=qwen|bonsai. When you pick Bonsai, seldon fetches its runtime + model (~7 GB, macOS/Apple-silicon) and runs the server as a managed daemon; seldon status shows the active brain, seldon down stops it.

Phone access: seldon up --tailnet serves the voice over HTTPS on your tailnet (via tailscale serve) and prints a https://<machine>.<tailnet>.ts.net URL — open it on your phone and the mic works (browsers require HTTPS for the microphone; plain http can't). The default seldon up stays loopback-only, and the Mac's own http://localhost:8770 already works (localhost is a secure context).

Needs HTTPS Certificates enabled for your tailnet — one toggle at login.tailscale.com/admin/dns → HTTPS Certificates; --tailnet tells you if it isn't. The serve config persists across reboots. Since Demerzel can act on the host, anyone on your tailnet can talk to it — use DEMERZEL_READONLY=1 when sharing.

seldon up only starts what's installed. demerzel serves the voice UI on http://localhost:8770. Its model set (~25 GB) is a one-time download — seldon install demerzel offers to fetch it right there ([y/N]), and seldon up offers again if it's still missing. Nothing that big is ever pulled silently. Logs live in ~/.seldon/run/.

Swap the in-process Qwen model: set DEMERZEL_LLM to any MLX-compatible HF repo before install/up — the fetch and the runtime both follow it (this only changes the Qwen brain; for Bonsai use --brain=bonsai above, which is a separate llama-server, not an MLX repo).

DEMERZEL_LLM=mlx-community/<some-mlx-repo> seldon up

Uninstall

seldon uninstall comb demerzel   # remove specific modules (asks to confirm)
seldon uninstall --all           # remove the whole stack (also drops ~/.seldon)
seldon uninstall <ids> --yes     # skip the confirm prompt

Node tools are removed from the global store; demerzel's isolated venv and bifrost's binary + config are deleted. Your system Python, tmux, tailscale, sops/age are never touched.

The stack

| module | what it is | ships via | |---|---|---| | foundation | the seam — a deterministic plan agents follow | npm | | apiary | the conversation — shared rooms where agents coordinate | npm | | comb | the vault — keys by name, never by value | npm | | factory | the floor — 24/7 supervisor + sandboxed agents | npm | | bifrost | the bridge — many machines, one workspace | shell | | demerzel | the voice — a fully-local voice assistant | python (macOS / Apple silicon) |

Ticking factory auto-includes apiary + foundation (it needs them on PATH).

Pick the combo for the job

It's a stack, not a bundle — install only what you use:

• Solo, one agent → seldon install foundation comb — a plan + secured keys. Add factory for the sandbox, demerzel for voice.
• A team of agents → add apiary (shared rooms) + factory (the supervisor).
• Across machines / from your phone → add bifrost.

Package managers

Node tools install with pnpm when it can global-install (much faster than npm), else npm. bun is supported but opt-in — seldon install --pm=bun (also --pm=pnpm|npm to force any of them).

demerzel (Python) always installs into its own isolated venv — never your system or active Python. It pins CPython 3.12 (its kokoro pin caps Python at <3.13) and uses uv to provision a standalone 3.12: if uv isn't present, seldon bootstraps it into ~/.seldon/bin with no shell/PATH changes. If uv can't be installed, it falls back to a matching python3.12 on PATH.

From a checkout

seldon --dev           # link Node modules from a monorepo checkout instead of npm

Part of the seldon monorepo.