wick-formatter 0.1.0

Compact text formats for feeding structured data into LLM contexts.

wick-formatter

Compact text formats for feeding structured data into LLM contexts. Python library, CLI, and stdio MCP server.

wick-formatter's primary format is LEAN (LLM-Efficient Adaptive Notation), an independent Python implementation of a format originally designed by Denys Fiialko. On tabular structured data, LEAN typically produces substantially fewer characters than uncompressed JSON while remaining losslessly round-trippable.

Guarantees

100% lossless round-trip. For any supported value:

decode(encode(data)) == data

This is verified on every benchmark run. If round-trip fails on any item, the benchmark hard-fails before measuring anything else.

No accuracy loss. LLM task accuracy with LEAN-encoded data matches JSON-encoded data within measurement noise. The benchmark enforces a maximum 3 percentage point delta.

Latest benchmark

Metric	Value
Round-trip	100% lossless
Compression	45.7% fewer characters
LLM accuracy (JSON)	53.8%
LLM accuracy (LEAN)	51.9%
Accuracy delta	1.9pp
Items tested	52
Model	llama3.1:8b-instruct-q4_K_M
Corpus	WikiTableQuestions + custom

See docs/BENCHMARK_RESULTS.md for methodology, data sources, and replication instructions.

Status

v0.1.0, pre-release. API and on-wire format are still stabilising. See docs/CHANGELOG.md for release notes and docs/ROADMAP.md for the v0.1.0 acceptance gates and deferred items.

Install

pip install wick-formatter

Runtime dependency: mcp (only required when using the MCP server).

What's in the box

• wick_formatter — Python library with a pluggable format registry.
• wick-formatter — CLI (--format=X {encode,decode}, stdin→stdout).
• python -m wick_formatter.mcp — stdio MCP server exposing wf_encode and wf_decode.
• tests/benchmarks/ — WikiTableQuestions-based harness comparing LEAN to a JSON baseline on llama3.1:8b-instruct-q4_K_M through an OpenAI-compatible endpoint (Ollama by default, API-provider swappable).

Quick example

Encode a small array of records into LEAN's tabular form:

echo '[{"a": 1, "b": 2}, {"a": 3, "b": 4}]' \
    | wick-formatter --format=lean encode

Decode it back:

wick-formatter --format=lean decode < record.lean

The full round-trip contract — including the ~-marker semi-tabular path, dot-flatten, and block encodings — is documented in docs/SPEC.md.

Python API

from wick_formatter import get, decode

# Get the LEAN format encoder
lean = get("lean")

# Encode tabular data
data = [{"name": "Alice", "score": 95}, {"name": "Bob", "score": 87}]
encoded = lean.encode(data)
print(encoded)
# name|score
# Alice|95
# Bob|87

# Decode back to original
decoded = decode(encoded)
assert decoded == data

MCP Server Setup

For Claude Code or Codex integration:

git clone https://github.com/p6rguvyrst/wick-formatter
cd wick-formatter
make client-claude   # or: make client-codex

Check status: make client-status

Remove: make client-clean

Requires jq for JSON manipulation. Install with brew install jq (macOS) or apt install jq (Linux).

Format specification

The complete LEAN format specification, including the encoder strategy selection rules and error cases, lives at docs/SPEC.md.

License and attribution

wick-formatter is released under the MIT License.

The LEAN format was originally designed and implemented by Denys Fiialko; see NOTICE for attribution, clean-room posture, and a specific credit for the semi-tabular encoding path first implemented in his toon-mcp-server repository.