Count tokens, truncate text to a token budget, and chunk text into budget-sized segments via @mukundakatta/promptbudget-mcp (npx)

@mukundakatta/promptbudget-mcp — token-budget-aware text truncation

Package: @mukundakatta/promptbudget-mcp (npm) Transport: stdio Install: npm install @mukundakatta/promptbudget-mcp → binary at node_modules/.bin/promptbudget-mcp Credential-free: yes (pure computation, no external calls)

Tools (3)

Key Gotchas

1. Token counting is APPROXIMATE — uses a simple ceil(text.length / chars_per_token) formula, default chars_per_token=4. This is NOT a real tokenizer (no BPE/SentencePiece). For Claude/GPT-4 the 4:1 ratio is a reasonable rough estimate but will be off for code, CJK text, or heavily punctuated text.
2. `strategy: "middle"` is BROKEN — returns an unparseable error. Only head and tail work reliably.
3. `count_tokens` with `max_tokens` returns `fits` boolean + `overflow_tokens` — useful for pre-flight checks before sending to an LLM.
4. `truncate_to_token_budget` cuts at character boundaries, NOT word/sentence boundaries — truncated output may break mid-word (e.g. "Foxt" from "Foxtrot").
5. Custom `marker` param is accepted but does NOT appear in output — the truncation marker (e.g. "... [TRUNCATED]") is consumed internally but not visible in the truncated field.
6. `chunk_to_budget` overlap is character-based — overlap_tokens: 2 means ~8 characters of overlap between adjacent chunks, not 2 semantic tokens.
7. First call is ~100ms (JIT), subsequent calls 2-25ms.
8. Custom `chars_per_token` lets you adjust for different tokenizers (e.g. chars_per_token: 2 for CJK text, chars_per_token: 6 for terse code).
9. `chunk_to_budget` returns a `chunks` array with chunk_count and total_tokens metadata — useful for planning RAG ingestion pipelines.
10. All tools return structured JSON — no plain text responses, easy to parse programmatically.