Skip to main content

hermes-token-speed

❖ Communityv1.1.0

Desktop · by Ayshinko · added

Live decode tokens-per-second chip for the Hermes Desktop status bar: 1-second sliding window while the model is still streaming, tool execution time excluded from decode seconds, UI spinner frames never counted, TTFT tracked separately, cumulative totals calibrated against authoritative provider usage at turn end. Desktop-only (desktop/plugin.js); no agent tools, hooks, middleware or env vars.

Open in Hermes Desktop
hermes plugins install hermes-token-speed
Repository ↗Reviewed source @ 0c0bb14 ↗Documentation ↗

README

From the reviewed commit 0c0bb14 ↗; it updates when the author re-pins.

hermes-token-speed

Live tokens-per-second for Hermes Agent desktop. A status-bar chip that shows your model's actual decode speed while it is still generating — not a post-turn average.

While the model streams: ⚡ 12.4 t/s (updates ~4×/sec). After the turn: ⚡ 13.9 t/s · 4210 tok · 302s (task-level cumulative, dsh-style).

Why this exists

Hermes desktop ships a built-in Tokens per second status field, but it is a rolling average over the last 10 completed API calls (see tui_gateway/server.py + agent/turn_usage.py). During one long generation its counters are frozen — the number you want most (how fast is it going right now) is exactly what it cannot show. Turn-level third-party meters (hook-based agents: Claude Code / Codex) can't fix this either: hooks only expose turn boundaries, so they can only report end-to-end speed with prefill and tool time baked in.

This plugin measures decode speed honestly, inside the desktop renderer's event stream:

Mechanism Detail
Sliding window 1000 ms window, burst-safe same-timestamp span extension (min 100 ms) so a coalesced chunk never fakes a spike
Tool time excluded tool.start / tool.generating pause the engine; decode seconds never include tool wall time
Spinner noise excluded thinking.delta frames are UI spinner rewrites, not model tokens — never counted
TTFT time from turn start to first streamed token, surfaced in turn report
Calibration mid-turn counts are regex estimates; at turn end authoritative provider usage (session.usage.output) reconciles all cumulative totals

It is a port of the measurement core from my Pi extensions pi-token-speed (sliding window) and pi-dsh-tps (task-level TPS excluding tool time) to the Hermes desktop plugin SDK. If you want the same numbers in other agents, see llm-speed-bench and openai/codex#54655 — this repo covers Hermes, honestly.

Install

  1. Create the folder <HERMES_HOME>/desktop-plugins/token-speed/
    • Windows: %LOCALAPPDATA%\hermes\desktop-plugins\token-speed\
    • macOS/Linux: your $HERMES_HOME/desktop-plugins/token-speed/
  2. Copy plugin.js into it.
  3. Done — standalone desktop plugins load by default. If you don't see the ⚡ t/s chip at the right end of the status bar: Ctrl+Shift+S toggles the bar, ⌘K → Reload desktop plugins re-scans.

Confirm the plugin is live: open the command palette (⌘K), type Token Speed — you should see turn report / watch events / reset. Chip missing but commands present → the status bar is hidden or the window is too narrow (right-end fields get trimmed first).

Commands (⌘K)

Command What it does
Token Speed: turn report toast: est tokens, decode seconds, LLM response count, TTFT, live TPS
Token Speed: watch events (5s) collect renderer event names for 5 s into a toast — use it after a Hermes update if the live number stops moving
Token Speed: reset zero the session chip totals

Tuning

Edit constants at the top of plugin.js (saves hot-reload in place):

WINDOW_MS   = 1000  // sliding-window width; raise for a calmer chip
MIN_SPAN    = 100   // burst coalescing floor; keep unless you see fake spikes
THROTTLE_MS = 250   // chip repaint-rate cap

Mid-turn token estimates use a word/punctuation regex (CJK = 1/char); expect ±20% drift vs. the provider tokenizer on mixed text — turn-end usage calibration removes the drift from all cumulative numbers.

Event contract (tested on Hermes desktop, Oct 2026)

From apps/desktop/src/app/session/hooks/use-message-stream/gateway-event/:

  • message.delta, reasoning.delta → real tokens, text in evt.payload.text
  • thinking.delta → UI spinner frames, must NOT count
  • tool.start / tool.generating / tool.complete → pause/resume wall time
  • session.usage → authoritative output for calibration

If a Hermes update renames these, watch events (5s) shows the real names in seconds. Note: a plugin that fails to load is not retried by the file watcher — fix + one manual ⌘K reload.

中文说明

Hermes 桌面版内置的 tok/s 是"最近 10 个已完成 API 调用的滚动平均"——一次长 生成进行中数字冻结，看不到实时速度。本插件把我在 Pi 上两个插件 （pi-token-speed / pi-dsh-tps）的测量核心移植到 Hermes 桌面插件 SDK： 1 秒滑动窗口实时读数、同时间戳分片防假尖峰、工具执行时间剔除、UI 转圈动画帧 （thinking.delta）不计数、回合结束用官方 usage 校准累计值。 安装：把 plugin.js 放进 %LOCALAPPDATA%\hermes\desktop-plugins\token-speed\， 状态栏右端出现 ⚡ 芯片；⌘K 搜 Token Speed 有三个命令。 Hermes 更新后若实时数字消失，跑 watch events (5s) 抓事件名改三行即可。

License

MIT — see LICENSE. Hermes is a trademark of Nous Research; this is an independent community plugin.

← Back to the catalog · catalog built Oct 20, 2026