MCP file tools silently eat your context window.I built one that doesnt

Implements file reading operations that track and report token consumption before returning content, using a token counter (likely tiktoken-based) to estimate context window impact. Unlike standard MCP file tools that silently consume context, this capability exposes token costs upfront, allowing clients to make informed decisions about whether to read files or use alternative strategies like summarization or chunking.

Provides file reading strategies that split large files into token-bounded chunks rather than returning entire files, using token counts to determine chunk boundaries instead of arbitrary line counts. The implementation likely uses a sliding window approach that respects semantic boundaries (e.g., function/class definitions) while staying within token budgets, allowing clients to incrementally load only the portions of files they need.

Maintains a session-level token budget that tracks cumulative consumption across multiple file read operations, enforcing limits before operations exceed the budget. The implementation likely uses a state machine or middleware pattern to intercept file tool calls, check remaining budget, and either allow, deny, or suggest alternative operations (like summarization) based on available tokens.

Calculates and returns token cost estimates for file operations before execution, using a tokenizer matched to the target LLM model. The implementation likely pre-tokenizes file content or uses heuristic estimation (characters × 1.3 for English text) to provide instant cost feedback without actually reading the file, enabling cost-benefit analysis before committing to expensive operations.

Implements directory listing and recursive file discovery operations that calculate and report cumulative token costs for all files in a directory tree. The implementation likely walks the file system, collects file metadata, estimates tokens for each file, and aggregates costs, allowing clients to understand the full token impact of loading an entire directory before committing to the operation.

Automatically selects and switches between tokenizers based on the target LLM model identifier, ensuring token estimates and counts match the actual model's tokenization scheme. The implementation likely maintains a registry of model-to-tokenizer mappings (e.g., gpt-4 → tiktoken, claude-3 → sentencepiece) and dynamically loads the appropriate tokenizer, with fallback heuristics for unknown models.