Forgehaven Labs
InferWatch

InferWatch · Features

Everything InferWatch 1.0 does.

Everything below ships in InferWatch 1.0 and was checked against the release build's source. Features marked Pro need the one-time unlock.

  • iPhone
  • iOS 17 or later
  • vLLM
  • Ollama
  • LiteLLM
  • No account
InferWatch app icon

Live telemetry

The numbers that tell you whether a rig is healthy right now.

  1. KV-cache pressure ring

    A ring you can read across the room, from vLLM's own cache-usage metric, on both the v0 and v1 metric names.

  2. Decode tok/s, TTFT and queue

    Decode tokens per second from vLLM's token counter, mean time to first token, and running and waiting requests on every vLLM rig.

  3. Refresh every 2, 3 or 5 seconds

    Pick the polling interval in Settings. Live polling runs while the app is open.

  4. Fleet dashboardPro

    Every rig on one screen, with rigs online and combined fleet tok/s up top.

  5. History sparklinesPro

    Recent tok/s and KV-cache sparklines on each vLLM rig. They hold about the last 60 readings and start over when the app restarts.

Your servers

Connect straight to the servers you already run.

  1. vLLM and Prometheus endpoints

    Reads vLLM's Prometheus /metrics. The Prometheus server type reads the same vLLM metric names from any /metrics endpoint that exposes them.

  2. Ollama models

    See which models are installed on each Ollama host, from /api/tags.

  3. LiteLLM gateway health

    Health, liveliness, database status and version for a LiteLLM gateway, with no key needed for the health check.

  4. LiteLLM token stats

    Token, request and latency totals when the gateway's Prometheus callback is on. Without it, the card shows health only.

  5. Home network and VPN

    Works with LAN addresses and VPNs like Tailscale. Allow Local Network access and InferWatch reaches homelab boxes directly.

Breach alerts

A local notification when something goes wrong, and another when it recovers.

  1. Four breach conditionsPro

    KV-cache above your ceiling (default 90 percent), queue backlog (default 20 waiting), generation stalled at 0 tok/s while serving, and server unreachable.

  2. Your thresholdsPro

    Set the KV-cache ceiling from 50 to 99 percent and the queue limit from 1 to 200, and switch the unreachable and stall alerts on or off.

  3. One alert per incidentPro

    No repeat nagging while a breach persists, and a Recovered notice when the rig comes back.

  4. Background checksPro

    InferWatch keeps checking when it is closed, as often as iOS allows: at most every 15 minutes, and often less. Best effort, not a pager.

Home Screen, demo and privacy

The rest of the kit.

  1. Home Screen widgetPro

    Rig status, tok/s and KV-cache in a small or medium widget.

  2. Demo rigs

    A labeled sample fleet (a vLLM rig, an Ollama Mac Studio and a LiteLLM gateway) to explore every screen. Demo rigs never use your free slot.

  3. Tokens in the Keychain

    Bearer tokens live in the iOS Keychain and are sent only to your server, only over HTTPS.

  4. Data Not Collected

    No account, no backend, no analytics, no tracking. InferWatch talks only to the servers you add.

Pricing

Free for one rig. Pay once for the fleet.

No subscription. No ads. InferWatch Pro is a one-time $14.99 unlock.

Free

$0

One real server, live.

  • One server with the live dashboard
  • tok/s, TTFT, queue depth, KV-cache
  • vLLM, Ollama and LiteLLM
  • Demo rigs, which don't use your free slot
  • Unlimited rigs and fleet view
  • Breach alerts
  • History charts and Home Screen widget
One-time

InferWatch Pro

$14.99

One-time purchase. No subscription.

  • Everything in Free
  • Unlimited rigs and the fleet view
  • Breach alerts (local notifications)
  • Recent history charts on every vLLM rig
  • Home Screen widget, small and medium
Get notified at launch

Price shown in USD; the App Store shows your local price.

Good to know

  • iPhone only, iOS 17 or later, portrait.
  • HTTPS servers need a certificate iOS trusts: InferWatch can't accept a self-signed certificate.
  • KV-cache, queue and stall alerts need vLLM-style metrics. Ollama and LiteLLM rigs get the unreachable alert.
  • There is no edit screen in 1.0: to change a rig, remove it and add it again.

Coming soon

InferWatch is heading to the App Store.

Drop a line and I'll tell you the day it lands.

Get notified at launch Read the docs

InferWatch is an independent app and is not affiliated with or endorsed by the vLLM, Ollama, or LiteLLM projects. Their names are used only to describe compatibility.