LLM Token Count

Real LLM token counts for GLM, GPT, Qwen & DeepSeek in Obsidian's status bar — so you know if a note fits the context budget before you send it.

The token count appears in the status bar alongside Obsidian's built-in word and character counts — live, for whatever note you have open.

• See token counts before you send. The status bar shows the active note's token count under your chosen model's tokenizer, so you can tell at a glance whether it fits the context window or needs chunking/summarizing.
• Accurate for GLM-5.2 — not a GPT approximation. The only Obsidian token-count plugin that tokenizes GLM-5.2 with its real HuggingFace tokenizer. GPT-based counters over-count Chinese by 10–30%, which misleads your context-budget decisions if you actually run GLM.
• 11 models, honestly labeled. GLM, GPT, Qwen, and DeepSeek are exact; Claude and Gemini are approx (o200k_base × 1.15, clearly labeled — never silently passed off as exact).
• Offline after first use. The tokenizer is fetched once from HuggingFace and cached locally. Subsequent counts are instant and need no network. GPT/Claude/Gemini need no fetch at all.

1. Download main.js, manifest.json, and styles.css.
2. Place them in your vault at .obsidian/plugins/llm-token-count/.
3. In Obsidian: Settings → Community plugins, reload the plugin list, enable LLM Token Count.

On first use with a HuggingFace-sourced model (GLM, Qwen, DeepSeek), the plugin fetches the tokenizer and caches it locally. The first count takes a few seconds; every count after is instant.

Open Settings → LLM Token Count:

• Default model — dropdown of all 11 supported models. Changing it re-counts the active note immediately.
• Re-download tokenizer — clears the cached tokenizer for the current model so the next count re-fetches from HuggingFace. Use after a tokenizer version bump or if the cache becomes corrupt.

• Obsidian 1.4.0+, desktop only. The plugin uses Node's fs and fetch for the tokenizer cache, which aren't available on mobile.
• Coexists with other token-count plugins (TokenBar, Token Count, TikToken Tokenizer, LLM Token Counter). This plugin's differentiator is GLM support and multi-model exact counting via HuggingFace tokenizers.
• No data leaves your machine except the one-time HuggingFace tokenizer fetch. Notes are never sent anywhere — tokenization is fully local.

GLM/Qwen/DeepSeek tokenizer SHAs are pinned to a specific HuggingFace commit for reproducible counts across machines and over time.

MIT