conjure-llm 0.0.1

Zero-dependency programming — replace library imports with LLM-generated, verified code

conjure

Zero-dependency programming — replace library imports with LLM-generated, verified code.

Every library you import is attack surface you didn't write. Conjure replaces thousands of transitive dependencies with one model binary and human-readable YAML specs.

Traditional app:
  App → pip install → 200 packages → 1500 transitive deps → 8M LOC of stranger code

Conjure app:
  App → .yaml spec files → embedded LLM → verified code → cached
  Trust chain: one model file (checksummed)

Install

pip install conjure

Quick Start

import conjure

# First call: generates, verifies, and caches (~10s)
result = conjure.invoke("base64_encode", data="hello")
print(result)  # "aGVsbG8="

# Second call: cache hit (0.3ms)
result = conjure.invoke("base64_encode", data="world")

How It Works

1. Write a YAML spec describing what you need:

spec: edit_distance
version: 1.0.0
description: |
  Compute the Levenshtein edit distance between two strings.
function: levenshtein
input:
  s1: str
  s2: str
output: int
examples:
  - input: { s1: "kitten", s2: "sitting" }
    output: 3
  - input: { s1: "", s2: "hello" }
    output: 5
constraints:
  no_imports: true
  max_lines: 400

2. Conjure generates a self-contained Python implementation (no imports), verifies it against your examples, and caches it:

result = conjure.invoke("levenshtein", s1="kitten", s2="sitting")
# Returns: 3

The generated code is:

• Import-free — enforced by AST analysis, not heuristics
• Verified — must pass all spec examples before caching
• Sandboxed — runs with restricted builtins and timeouts
• Cached — content-addressed by spec + model + seed (sub-ms on repeat calls)

Results

On ConjureEval-100 (100 specs across 20 categories):

Metric	Rate
pass@1 (first attempt)	70.0%
pass@3 (best of 3 attempts)	87.9%

The retry pipeline recovers ~18% more specs by giving the model its own error messages.

Attack Surface Reduction

Across 5 real Python applications:

Application	Dependencies	Conjure	Reduction
Flask blog	13 transitive	0	15× LOC
FastAPI service	15 transitive	0	17× LOC
CLI tool	5 transitive	0	6× LOC
Web scraper	17 transitive	0	20× LOC
File sync	8 transitive	0	9× LOC

Translation Feasibility

52% of a typical Flask application's imports can be automatically replaced by Conjure specs.

CLI

# Invoke a function
conjure invoke base64_encode -k data=hello

# Pre-generate all specs
conjure build --spec-dir specs/stdlib

# Evaluate specs
conjure eval --spec-dir specs/stdlib --output results.md

# Analyze a project for translation
conjure translate my_project/ --output conjure_specs/

# Compare attack surface
conjure audit --packages "flask,requests,pyyaml" --specs specs/stdlib

Model

Conjure uses Qwen3.5-9B-OptiQ-4bit via MLX on Apple Silicon:

• 5 GB model memory (4-bit mixed-precision quantization)
• Instruct mode with recommended sampling (temp=0.6, top_p=0.95)
• Runs entirely on-device — no API keys, no cloud, no network

Included Specs

30 curated stdlib specs covering common library patterns:

Category	Specs
Encoding	base64_encode, base64_decode, hex_encode, hex_decode, rot13
String	slugify, camel_to_snake, snake_to_camel, reverse_words
Collections	group_by, flatten, flatten_dict, chunk, deduplicate
Data parsing	csv_parse, json_parse, url_parse
Math	gcd, fibonacci, statistics, levenshtein
Algorithms	binary_search, run_length_encode, matrix_multiply
Other	is_palindrome, word_count, truncate, frequency_count, deep_get

Plus 970 additional specs across 20 categories available in the full distribution.

Limitations

• Complex algorithms: SHA-256, recursive-descent parsers — too complex for reliable generation from a 9B model
• Apple Silicon only: Requires MLX (macOS with M-series chip)
• Cold start: First generation takes 3-100s (cached permanently after)
• Pure functions only: Stateful protocols, FFI, OS access are out of scope

License

MIT