Token Speed
Live token generation speed and session up/down token totals in the Hermes Desktop status bar — plus a click-through dashboard.

⚡ 38.4 t/s ↑12.4k ↓3.2k sits in the status bar while you work. Click it for the full picture:

What it shows
| Surface | Values |
|---|---|
| Status bar chip | current tokens/sec, ↑ prompt tokens used this session, ↓ completion tokens this session |
| Dashboard (click the chip) | 60-second sparkline, this-turn output tokens, model, call count, average t/s, cache hit %, context usage %, estimated cost (USD) |
| Command palette | Token Speed: reset samples — clears the sliding window and history |
Speed reads 0 and the chip goes quiet when nothing is streaming. No spinner, no flicker.
Install
Catalog (once listed):
hermes plugins install token-speed
From source:
hermes plugins install https://github.com/seagull/hermes-token-speed
Manual, desktop-only (works without any agent-side plugin machinery): copy desktop/plugin.js to
$HERMES_HOME/desktop-plugins/token-speed/plugin.js and restart Desktop. That directory is app-wide, so every
profile sees the same copy.
Then enable it in Settings → Plugins.
How the number is produced
The plugin never guesses from a clock alone and never asks the model for a rate:
message.delta/reasoning.deltaframes feed a 2-second sliding window of character counts
(thinking.deltais deliberately excluded — the core rewrites it with spinner frames, not real output).- A 250 ms tick converts chars → tokens with a learned chars-per-token ratio.
UsageStatsticks (authoritativeinput/output) recalibrate that ratio continuously
(exponential smoothing, 0.65/0.35), clamped to 0.7–12 chars/token so CJK-heavy (~1.2) through
English code (~4.5) both track and a miscount can't poison it.- Between frames the speed decays geometrically instead of snapping to zero mid-sentence.
↑ and ↓ are read straight from the session's UsageStats — no estimation at all.
Privacy and footprint
- No network calls. No telemetry. Nothing leaves the machine.
- No file writes, no credential reads, no env vars (
requires_env: []). - Only the Desktop plugin SDK (
@hermes/plugin-sdk,react/jsx-runtime); no DOM scraping, no
prototype patching, noeval, no core overrides. - All state is module-scoped and resets when the app reloads; counters follow the session.
Compatibility
- Hermes Desktop >= 0.23.0 (
requires_hermes). - Desktop-only: no agent tools, hooks, middleware or provider surfaces. The manifest declares empty
capabilitieson purpose — validation fails an entry whose declarations drift from what it registers.
Layout
plugin.yaml manifest (v2), desktop-only
desktop/plugin.js the whole plugin: plain ESM, loaded uncompiled (jsx() calls, not JSX syntax)
assets/ screenshots
中文说明
状态栏实时显示 token 生成速度 和 本会话上下行总数,点一下看完整面板。
| 位置 | 内容 |
|---|---|
| 状态栏小条 | 当前 tok/s、↑ 上行(prompt)、↓ 下行(completion) |
| 点击弹出的面板 | 60 秒柱状图、本轮输出 token、模型、调用次数、均速、缓存命中率、上下文占用、预估成本 |
| 命令面板 | Token Speed: 清空采样 |
速度靠 2 秒滑动窗口的字符流估算,并用 UsageStats 的真实 input/output 持续自校准字符/token 比例
(中英混排都跟得住);不联网、不写文件、不读凭证,只用插件 SDK。
安装:hermes plugins install token-speed,或在 设置 → 插件 里启用。
License
MIT
README from the reviewed commit 5c210d4; it updates when the author re-pins.
shengqiqwq-spec