Smart token compression MCP server for LLM apps. Save 70-90% on API costs.
PromptThrift provides 4 tools with complete input schemas and descriptions. Naming follows verb_noun convention (compress_history, count_tokens, suggest_model, pin_facts). Descriptions are present and substantive (100-150 chars each). However, output schemas are not documented, LLMs cannot predict return structure. Parameter descriptions are adequate but lack format constraints (e.g., no enum for pin_facts 'action' param). Error handling is minimal; no recovery guidance. Security is sound (no secrets in params, local processing). Tool composition is reasonable but compress_history and count_tokens overlap in purpose (both optimize tokens). Missing: pagination guidance, per-item error reporting, and actionable error messages.
Summarize old conversation turns to reduce input tokens
Track token usage and estimate costs in real-time
Pin critical facts that must survive compression
Recommend the cheapest sufficient model for a task
Output schemas not documented. LLMs cannot predict return structure (fields, types, pagination). compress_history, count_tokens, suggest_model, and pin_facts all lack documented return types.
pin_facts 'action' parameter lacks enum constraint. Description says 'add, remove, or list' but no formal enum in schema. LLMs may hallucinate invalid actions.
No error handling guidance. Tools do not document what errors can occur, whether they are retryable, or what the LLM should do next (e.g., if Ollama is unavailable, if model pricing is unknown).
compress_history and count_tokens both optimize token usage but serve different purposes. Tool composition could be clearer: compress_history modifies conversation; count_tokens only estimates. No guidance on when to use each.
Inferred effective spec: <=2025-11-25.
| Scored | Grade | Overall | Spec posture | Rubric |
|---|---|---|---|---|
| 2026-09-23 | F | 49 | <=2025-11-25 | v2 |
suggest_model parameter 'prefer_local' defaults to false but description says 'default: false'. No guidance on when LLM should override this default or how to interpret the recommendation.