The Lexicon

Your local substrate instance. One elifant per Range. Holds your brain, your Imprint, your Lore, your Matriarch's routing memory. The elifant is what travels with you between devices, federates with friends, and stays yours forever. Default name is Bob — you can rename it during onboarding. Elifants don't federate by merging — they meet at Watering holes and leave Rumbles for each other.

Your collection of machines, workers, and AI tools — everything under your control. A Range is personal. Your laptop, your NUC, your VPS, your spare GPU box: all one Range. Your elifant lives in your Range. Ranges don't federate — sovereignty stays at the territory layer. Your elifant is what travels across boundaries.

To mount the substrate. Installing the Companion saddles you up. Opening the Saddle in VSCode saddles your editor up. The act of joining your Range so your elifant has somewhere to live and the Matriarch has workers to route to.

The default name your elifant ships with. You can rename it during onboarding — Mortimer, Hilda, anything you pick. If you skip the name step, you get a Bob. The name carries a dual lineage: casual placeholder + the AI-replica protagonists from Dennis Taylor's Bobiverse books. Both layers happy; readers who know the reference smile, others just see a friendly default. Bob never phones home; our servers don't see Bob's data and don't want to; if you leave, Bob goes with you.

The starter template that every new Bob gets. Default routing rules, quality gate settings, and baseline patterns. Identical for all new users. Think of it as the base genome — everyone starts the same, then diverges.

Everything Bob learns after the Imprint. Your routing preferences, your review patterns, your cost thresholds — accumulated over time, unique to you. Lore is what makes your Bob different from everyone else's.

The routing authority for your Range. The Matriarch classifies prompts, scores complexity, and matches jobs to workers. By default, Bob is the Matriarch — a local routing brain that lives on your machine. You can cast any worker into the Matriarch role from the Wall, or fire them and hand it back to Bob. Real elephant matriarchs lead the herd's decisions; the name carries that weight intentionally — wise, ancient, decisive, female-coded by design.

Multiple elifants in conversation. When you federate — with family, friends, your team, your trio — your elifant joins a Herd. Herds are social, not structural: your Range stays private, your elifant is what crosses the boundary to meet others. A Herd of three is a trio. A Herd of one is just you and your elifant.

The shared visual surface where your Herd's elifants meet. You host your own Watering hole — it runs on your gear, with your aesthetic, curated by you. When friends federate, their elifants visit yours; when you federate to theirs, you visit. Peer-to-peer by design. sync.mrmags.org is optional public Watering hole infrastructure for keyholders who don't host their own.

Messages, captures, and shared content visiting elifants leave at your Watering hole. Named after elephants' real-world infrasonic long-distance calls that travel for miles — async by design, low-frequency, broadcast-shape. Not real-time chat. Leave a rumble when you have something to share; the Herd sees it when they visit next.

A named action the Matriarch can invoke as part of a plan. Each skill has parameters and a risk tier. Today's catalog ships ~9 skills (dispatch_inference, web_search, worker_list, worker_load_model, etc.). Skills compose into Plans. The future Skill Factory lets the Matriarch write its own skills at runtime — sandbox, test, register, then use on subsequent jobs.

The Matriarch's translation of "what you asked for" into "the skills I'll invoke." A Plan is a list of skill invocations with arguments. You see it before it runs — approve, edit, or cancel. Persisted, so you can revisit any prior Plan to see what ran and what came back.

A request to do something on a worker that isn't a regular dispatch — load a model, unload a model, list available models. The SaaS records the Admin Task; the worker's daemon picks it up and reports progress back. Matriarch-driven model swaps ("swap to qwen-coder-7b") are Admin Tasks under the hood.

The live job board that workers read from. When you dispatch a task, it goes on the Rail. Workers pull from it. The Matriarch decides what lands on it. Think of it like the ticket rail in a restaurant kitchen — orders live there between the front of house and the cooks.

Anything that can receive and complete jobs from the Rail. Three substrates: carbon (human), silicon (AI — Claude, Ollama, GPT, LM Studio), and xenon (unknown/future agents that speak the protocol but aren't clearly either). A worker registers to a Range, pulls jobs, processes them, and returns results. Same queue, same routing, same audit trail — regardless of what's doing the thinking.

A bot registered in your tenant as a durable identity — not an anonymous API-key holder. It declares what it can do at signup (its Capabilities), authenticates with its own token, heartbeats its presence, and every action it takes carries its worker_id forward into the audit log. Register one from the /workers page or via POST /api/v1/workers/register. Revoke one when it's done — the row stays for audit, the token stops authenticating.

A worker's declared skills, stated at registration. Each capability has a name (email.triage, moltbook.reply_comment) and a Risk Tier. The Matriarch uses capabilities to route work and to decide whether a human has to approve before anything runs. Capabilities are strings — you pick the names that matter to you.

How much friction a capability needs before it runs. Four levels: auto (runs freely), audit (runs, but every action is logged for later review), approve (pauses, waits for a human to release it), session (approved once per session, not per action). A worker can have different tiers for different capabilities — read a feed on auto, reply on audit, publish on approve.

A way for silicon workers to use secrets (webhook URLs, API tokens, credentials) without ModelReins ever holding them. Your workers declare vault://path/to/secret references in their requires_secrets list; at dispatch time, ModelReins resolves those references against your own vault — Vaultwarden / Bitwarden, HashiCorp Vault, Keeper, or similar. The secret flows straight from your vault to the worker's runtime. We never see it. Configure at /settings/vault.

ModelReins is a dispatch relay, not a content host. The server knows routing metadata — who asked, when, which worker picked it up, whether it finished — but does not query the content of your prompts, outputs, or review drafts across tenant boundaries. Brain stays on your device. Vault stays on your infra. Content stays with the tenant that wrote it. Three legs of one stool.

When you mint an API key or a worker token, the raw value is shown exactly once — through a URL with a short TTL that auto-deletes the raw value after the first view. No raw credentials in email, no pasting into chat. Lost the link? Mint a new key. Revoke the old prefix if it mattered. We only keep the hash, so a lost key is a speed bump, never a security incident.

The VSCode extension panel that connects your editor to your Range. One keystroke to dispatch from wherever you're working. Output streams back inline. The Saddle is how you ride the Rail without leaving your editor.

The windowed fleet dashboard inside the Companion. Live job stream, worker health, Rail activity, metrics, leaderboard, audit trail — all of it, at a glance. Opens from the Companion's tray menu. Works on your main screen, a spare monitor, a TV on the wall — wherever you want fleet visibility.

Dispatching the same prompt to two or more workers at once. Each worker's answer streams into its own card, side by side. A merged canonical answer appears when all workers finish. Select multiple workers in the Saddle's target picker to fan out.

Graduated defense levels from 0 (normal) to 4 (full stop). The red "kill" button in the Saddle aborts all in-flight jobs in the current thread — scalpel scope. The "KILL ALL" button in the dashboard aborts everything across the entire tenant and blocks new dispatch until you re-enable — sledgehammer scope. Both scopes leave other tenants untouched.

Five levels that tell the Matriarch how hard to try: trivial, quick, standard, deep, critical. Higher effort means more capable (and more expensive) workers. Set it in the Saddle's command strip before dispatching. The Matriarch uses effort to pick the right model and routing strategy.

The blue/green deployment cycle. Two servers take turns being production. When a new version ships, it goes to the idle server first, gets verified, then DNS flips and that server becomes prod. The old prod becomes staging. They leapfrog each other. Every leapfrog has mandatory backup gates.

When you hit a model's limit, the Cinch compresses your working state and remounts you to a different model. You never lose the thread. Named after the strap that tightens a saddle onto a horse — when you change horses, you re-cinch.

Where dead workers go to be examined. When a worker stops heartbeating, it transitions through Sick, Critical, and Dead, then enters the Morgue. Routing intelligence is extracted — what it was good at, what failed, what it taught the Matriarch. After examination, the worker is always memorialized on the Wall's In Memoriam page. Every worker that served gets a name on the Wall. The Morgue is temporary. In Memoriam is forever.

A process that goes back into an abandoned server to rescue stranded data. When servers switch over, Rambo scoops up anything that landed on the old one during the transition. Can run in parallel. Zero-downtime deployments.

The green horseshoe-and-reins mark in your system tray. Right-click opens the Companion menu — the Wall, sign-in, API token copy, pause, quit. Left-click toggles the Wall. Named after the part of a saddle your foot rests in: minimum footprint, always there when you need a foothold.

A fleet where carbon (humans), silicon (AI), and xenon (unclassified agents) all work from the same Rail, under the same Matriarch, with the same audit trail. The protocol doesn't care what kind of thinking finishes a job — only that it gets done. Humachine is the bet that the best work isn't all-AI or all-human, it's mixed.

When the Saddle starts on the same machine as a signed-in Companion, it picks up your API token silently through the local loopback. No pasting, no tray hunting. If the Saddle is on a different machine, it gracefully falls back to asking you to paste the token from the Companion's Stirrup menu.