Every Arth component runs in two shapes: single-user (everything on your machine, zero external services) and team/central (a shared server your whole team points at). This page is the map.
Availability: Local is GA for every layer. Per-layer cloud endpoints (telemetry ingest at
ingest.getarth.ai, your org’s own arth-memory server) work today; the unified Arth Cloud sign-in (one login for intelligence + memory + router + orchestrate) is pre-GA;/arth-setupsays so when you pick cloud.
Each layer is optional and independent. Pick any row in any combination, and upgrade a layer from single-user to central later without touching the others.
| Layer | What it gives you | Single-user shape | Team/central shape |
|---|---|---|---|
| Toolkit (arthai plugins) | Skills, agents, hooks, /calibrate knowledge |
Just the plugin: no services at all | Same plugin per seat; shared conventions via your repos |
| Arth Intelligence (experimental preview) | Sessions / cost / DAG dashboard (OTEL) | Local Docker stack via /otel-setup → Local (dashboard on localhost) |
Cloud endpoint (/otel-setup → Cloud, app.getarth.ai) or a self-hosted central engine |
| arth-memory | Semantic knowledge retrieval + constraint checks | /memory-setup → Local tier (Docker store on this machine; no discrete GPU required — on macOS Ollama runs natively and uses Metal when available) |
/memory-setup → Cloud (direct to org server) or Remote (offline replica of the org server) |
| decisions (Jev) | Typed, calibrated gates for /qa, /pr, /ci-fix and the router — shadow first, enforce per gate after calibration (guide) |
The decisions checkbox in /arth-setup; a hosted TypeSafe key in ~/.arthai/.env, or the local runtime shadow-only |
Same checkbox; central posture points the wrapper at the deployment proxy, no per-user key |
| arth-router | Model/harness routing with cost governance | Skip it: solo routing is the toolkit’s built-in triage; the router’s value (budgets, pools, org policy, decision audit) is org-shaped | Central router service; the toolkit’s triage hook calls it via the fast-path integration |
One command drives it all after the plugin install:
/arth-setup # telemetry? extras? knowledge backend? — 3 questions
/calibrate # per repo — generates the knowledge everything serves
files (zero infra); switch to the Local memory tier when
you want semantic recall or work across multiple repos. /memory-setup off
reverts instantly; the markdown knowledge base is the source of truth in
both modes.Hardware floor for the whole single-user stack: any Mac/Linux dev machine, Docker, no discrete GPU required — on macOS Ollama runs natively and uses Metal when available; without it, memory recall degrades to keyword search. Disk: ~1.5GB for Intelligence images + ~800MB for the Memory light path (store + embedding model), ~2.5GB if you take everything.
That floor covers the toolkit’s own Intelligence and Memory layers. Where your AI models run is a separate choice, with its own hardware floor — see Prerequisites for the short per-preset CPU/RAM/disk/GPU reference, or the Single-User Guide for the full walkthrough (setup questions, task division with a cheaper model, keys, feature switches).
Coming with the mixed-execution release: /arth-setup will ask a
second question right after posture — posture (Local or Central — is
this just you, or a team pointing at shared services?), then “Where should
AI models run?” with three presets: Claude only (default), Claude + one
cloud provider, or Claude + one local runtime. That second question drives
the internal provider-setup step: for a cloud provider it lists the models
whose tool-use probe passed, and for a vendor without a key it prints one
command for you to run in your own terminal, never in the chat:
bash ~/.claude-agents/scripts/arth-keys.sh set FIREWORKS_API_KEY # symlink install
bash "$CLAUDE_PLUGIN_ROOT/scripts/arth-keys.sh" set FIREWORKS_API_KEY # plugin install
The helper prompts with hidden input and stores the key in ~/.arthai/.env
(mode 600); arth-keys.sh wizard takes several keys in one sitting (TypeSafe,
the Cloudflare pair, Vercel, Fireworks) and --help says where each is
created. Setup then asks Done / Skip for now, detects the key by name
and runs the conformance probe (✓ tool-use verified or ✗ one-shot only).
The same helper takes every auth_env a provider descriptor names
(DEEPSEEK_API_KEY, DASHSCOPE_API_KEY, MOONSHOT_API_KEY, GLM_API_KEY,
and the decisions keys); list shows the names stored, unset NAME removes
one. With Claude only, every model call stays on Claude.
/otel-setup → Cloud/Custom writes it globally). Dashboards aggregate all
seats; nothing else on the laptop changes./memory-setup (Cloud / Local / Remote) and recorded in
~/.arthai/config.json; ARTH_MEMORY=off|local|cloud overrides it per run
for A/B tests. Skills are agnostic: the kg-query dispatcher asks the
shared arth-memory.sh resolver for the active tier and falls back to the
lexical graph on any failure. RBAC (reader ⊂ writer ⊂ admin) and per-scope
tokens apply on the server side./router-setup per seat
wires the env vars and the triage fast-path. Decisions, budgets, and
outcome telemetry aggregate org-wide.Full walkthrough — the four values an admin hands out, per-seat setup choices for each skill, and what stays admin-owned and seat-immutable — is in the Team Setup Guide.
Common real-world combinations:
| You are | Telemetry | Memory | Router |
|---|---|---|---|
| Solo dev | Local | files or memory-local | skip |
| Solo dev, multiple machines | Cloud | central + replicas | skip |
| Small team | Cloud or self-hosted central | central + replicas | optional |
| Org with cost governance | central | central | central |
files → memory-local → central never regenerates or deletes
knowledge, the calibrate markdown moves with you (import up,
hydrate down).