MCP Tools Reference

Distill provides 19 tools for optimizing LLM token usage. Each tool is designed for specific use cases.

File Reading Tools

smart_file_read

Read code files with AST-aware extraction. Returns structure overview or specific elements.

smart_file_read filePath="src/server.ts"
smart_file_read filePath="src/server.ts" target={"type":"function","name":"createServer"}
smart_file_read filePath="src/utils.ts" query="parse"
smart_file_read filePath="src/server.ts" skeleton=true

Savings: 50-70% compared to reading full files

Supported languages: TypeScript, JavaScript, Python, Go, Rust, PHP, Swift

code_skeleton

Extract only function/class signatures without implementation bodies.

code_skeleton filePath="src/api.ts" depth=2

Savings: 70-90%

Compression Tools

auto_optimize

Automatically detect content type and apply optimal compression.

auto_optimize content="<build output or logs>"

Savings: 40-95% depending on content

compress_context

Generic compression for any text content.

compress_context content="<large text>" targetRatio=0.5

Savings: 40-60%

semantic_compress

TF-IDF based compression that keeps important segments.

semantic_compress content="<document>" targetRatio=0.5

Savings: 40-60%

diff_compress

Compress git diff output while preserving essential changes.

diff_compress diff="<git diff output>" strategy="hunks-only"

Strategies: hunks-only, summary, semantic

Savings: 50-95%

conversation_compress

Compress conversation history while preserving key information.

conversation_compress messages=[...] strategy="hybrid" maxTokens=1000

Strategies: rolling-summary, key-extraction, hybrid

Savings: 40-70%

Build & Error Tools

analyze_build_output

Parse and compress build errors from npm, tsc, webpack, etc.

analyze_build_output output="<build output>" buildTool="tsc"

Savings: 95%+

deduplicate_errors

Group repeated errors and show occurrence counts.

deduplicate_errors content="<output with repeated errors>"

Savings: 80-95%

summarize_logs

Summarize verbose logs (server, test, build, application).

summarize_logs logs="<log content>" focus=["errors","warnings"]

Savings: 80-90%

detect_retry_loop

Detect if Claude is stuck in a retry loop.

detect_retry_loop command="npm run build"

Analysis Tools

analyze_context

Analyze content for token usage and get optimization suggestions.

analyze_context content="<prompt or context>"

context_budget

Pre-flight token estimation before sending to LLM.

context_budget content="<content>" budgetTokens=5000

session_stats

View current session statistics.

session_stats detail="detailed"

get_stats

Get usage statistics for the session.

get_stats period="session"

optimization_tips

Get context engineering best practices.

optimization_tips focus="prompts"

Discovery Tools

discover_tools

Find and load optimization tools on-demand. Supports TOON format for compact output.

discover_tools category="compress" load=true
discover_tools query="logs"
discover_tools format="toon"
discover_tools format="toon-tabular"

Categories: compress, analyze, logs, code, pipeline

Formats:

• list - Default human-readable format
• toon - TOON format grouped by category (55% fewer tokens)
• toon-tabular - TOON tabular format (most compact)

TOON Output Example:

tools[15]:
  auto_optimize(content hint?:build|logs|... aggressive?:bool) → Auto-compress 80-95%
  smart_file_read(filePath target?:{type,name} query?) → AST code extraction
  compress_context(content contentType?:logs|...) → Compress verbose text 40-90%
  ...

[tokens] json:1189 → toon:531 (-55%)

Utility Tools

smart_cache

Interact with the file cache for parsed content.

smart_cache action="stats"
smart_cache action="get" key="file:/path/to/file.ts"

smart_pipeline

Chain multiple compression tools automatically.

smart_pipeline content="<mixed content>" mode="auto"

register_model

Register the Claude model being used for accurate cost tracking.

register_model model="claude-opus-4-5-20251101"

Token Savings Summary

About TOON Format

TOON (Token-Oriented Object Notation) is a compact, human-readable format designed specifically for LLM inputs. It provides:

• 55% fewer tokens compared to JSON for tool listings
• Lossless representation of structured data
• Human-readable syntax with YAML-like indentation

TOON is particularly effective for tabular data and uniform arrays, making it ideal for presenting tool catalogs to LLMs.