Ulanzi Inference Monitor

Live usage limits and balances for your inference providers, drawn on your Ulanzi deck keys — a sibling of ulanzi-system-monitor that reuses the same interface model (a cycler key + tiles that follow the active selection).

Latest release: v1.5.1 ships both the Ulanzi Studio plugin and the standalone inf-agent.

One key cycles the provider; two tiles show its main numbers, with an optional third monthly tile for providers that expose it:

Provider	Kind	Primary tile	Secondary tile
Claude (Anthropic)	limits	Session (5h) % + reset	Week (7d) % + reset
OpenAI (Codex)	limits	Short window % or plan	Long window % + reset
Ollama Cloud	limits	Session (5h) %	Week (7d) %
OpenCode Go	limits	Rolling 5h % + reset	Week % + reset (monthly on a third tile)
OpenRouter	balance	Balance ($)	Spend today / week
Nous	balance	Free / tier	Rate limit (rpm/tpm)

┌──────────┐  ┌──────────┐  ┌──────────┐
│ Provider │  │  ◐ 8%    │  │  ◔ 2%    │
│  Switch  │  │ SESSION  │  │  WEEK    │     ← Claude selected
│ 🤖 Claude│  │ ↻ 4h 15m │  │ ↻ 5d 23h │
│   1/6 8% │  └──────────┘  └──────────┘
└──────────┘   press switch →  $8.29 / today $0.02   ← OpenRouter selected

Architecture

 host with your keys (hermes NUC)              machine running Ulanzi Studio
┌───────────────────────────────┐            ┌──────────────────────────────┐
│ inf-agent.py                   │  HTTP      │ Inference Monitor plugin       │
│  reads .claude/.codex/.hermes  │  (Tailscale)│  ProviderSampler → /providers │
│  GET /providers  → all providers├───────────▶│  Switch key cycles providers  │
│  refreshes every 60s           │            │  Tiles render the active one  │
└───────────────────────────────┘            └──────────────────────────────┘

The agent does all provider-specific work and lives where the credentials already are; the plugin just polls one endpoint and renders. See agent/README.md for the JSON contract and per-provider sources.

Setup

1. Run the agent on the host with your keys

cd agent
python3 inf-agent.py            # listens on 0.0.0.0:9890
# or install as a service — see agent/README.md

Verify: curl http://localhost:9890/providers | python3 -m json.tool

2. Install the plugin in Ulanzi Studio

./pack.sh                       # builds dist/com.ulanzi.infmonitor.ulanziPlugin-<ver>.zip

Double-click the .zip (or import it in Ulanzi Studio). Then on the deck:

1. Drop a Provider Switch key. In its settings put the agent address — http://127.0.0.1:9890 if Studio runs on the same box, otherwise the agent host's Tailscale address, e.g. http://100.x.y.z:9890.
2. Drop two Provider Tile keys next to it; set one to Primary and one to Secondary.
3. In the switch settings, tick only the providers that this key should cycle. Selected rows appear under Cycle order; use ↑/↓ to set the exact order. Unselected rows remain under Available providers, and All / reset order restores every provider in the agent's order. Older switch settings continue to include all providers. Press the key to cycle the chosen providers in the saved order; the tiles follow the selection.

If OpenAI was absent from the cycle after installing v1.5.0, it was only omitted from that switch's saved allow-list — the agent still returned it. In v1.5.1 it remains visible under Available providers; tick OpenAI or use All / reset order to restore it.

OpenCode Go exposes three windows. Add a third Provider Tile and set it to Monthly to show its monthly limit; that slot also shows OpenRouter's monthly spend when available.

The Provider Switch owns the shared agent address when one is configured. Provider Tiles follow it, even if an older tile still contains the localhost default. Without a configured switch, a standalone tile uses its own address. The address is persisted for startup without opening the Property Inspector.

Local startup hotfix — v1.5.2, 2026-10-10 (software verified)

The installed Windows Studio 3.3.9 plugin uses the configured Provider Switch as owner of the shared agent endpoint; a restored tile's localhost default cannot override it. Fresh-process startup tests and the approved final software checks passed. A passive capture at 19:11:44–19:12:04 UTC observed five automatic GET /providers requests from Windows, each answered with HTTP 200. The installed app hash still matches the tested source. No URL/input interaction was used for these checks. [Source: final test output, Windows read-back and bounded passive HTTP capture, 2026-10-10]

Independent read-only review approved the hotfix. Physical-panel appearance was not observed remotely. The live preview retains a separate failed Ollama Cloud live-limits assertion; upstream provider errors remain visible, not replaced with invented readings. v1.5.2 is a local plugin build/source hotfix, not a new published GitHub release asset. See evidence and limitations. [Source: completed review, live preview and release scope, 2026-10-10]

Notes & limitations

• Claude session/week come from the Anthropic OAuth usage endpoint using the token Claude Code keeps fresh; the agent only reads it (never refreshes, so it can't disturb your Claude Code login). On token expiry the tile shows an error until Claude Code refreshes.
• OpenAI usage comes from the same read-only ChatGPT usage endpoint used by Codex (/backend-api/wham/usage). The agent reads /root/.codex/auth.json but never refreshes or writes it. Window labels follow the durations OpenAI actually returns (for example 5H and WEEK); accounts with only a weekly window show the plan on the first tile and the weekly gauge on the second. If the endpoint is briefly unavailable, the newest local Codex rate-limit snapshot is shown as stale until live polling recovers.
• Ollama Cloud exposes live session and weekly fractions at its authenticated GET /api/usage endpoint. The endpoint does not include reset timestamps, so its gauges intentionally do not invent a countdown. Plan/renewal metadata still comes from POST /api/me when available.
• OpenCode Go reads its key (prefer OPENCODE_GO_API_KEY, otherwise OpenCode's local auth.json) and uses GET /zen/go/v1/usage for rolling 5-hour, weekly, and monthly percentages and reset times. A rejected key or missing subscription is isolated to that provider and does not interrupt the other tiles.
• Nous rate limits/tier come from the portal JWT (no network call). Real plan + purchased balance are read live from the portal account API by delegating to hermes's own get_nous_portal_account_info() (run in the hermes venv) — hermes owns the single-use token refresh + persistence + locking, so the agent never calls the Nous refresh endpoint itself (reuse revokes the whole session). Configure via INF_NOUS_HELPER_PY / INF_NOUS_HELPER_CWD (defaults point at this box's hermes), cached INF_NOUS_LIVE_TTL (60s). If hermes has no Nous session, it falls back to the INF_NOUS_PLAN / INF_NOUS_BALANCE env values; restore live with hermes auth add nous.
• Dialagram is intentionally not included (no balance API; subscription inactive). Add providers by extending agent/inf-agent.py (PROBES).
• Secure the agent with INF_AGENT_TOKEN and/or bind it to your Tailscale IP if the host is reachable beyond the tailnet.

Layout

agent/inf-agent.py         provider probes + HTTP endpoint (stdlib only)
agent/inf-agent.service    systemd unit
com.ulanzi.infmonitor.ulanziPlugin/
  plugin/app.js            main loop: sampler + cycler + paint
  plugin/monitor/ProviderSampler.js   polls /providers (no-wedge fetch)
  plugin/monitor/render.js            ring-gauge & value-card tiles + switch tile
  plugin/monitor/settings.js          PI settings
  property-inspector/      tile + switch config UIs
pack.sh                    build the plugin + agent zips into dist/

Reuses the device API (plugin/common-node), PI framework (libs/) and MDI icon subset from ulanzi-system-monitor.