Zero-dependency programming
Replace imports
with specifications.
A local LLM generates self-contained Python from YAML specs. Verified against examples. Cached forever. Zero imports.
$ pip install conjure-llm
How it works
Spec in, verified code out
Write a YAML spec with a function signature and examples. Conjure generates, verifies, and caches the implementation.
levenshtein.yaml spec
name: levenshtein version: "1.0" function_name: levenshtein params: s1: str s2: str return_type: int examples: - input: {s1: "kitten", s2: "sitting"} output: 3
app.py usage
import conjure # First call: generate, verify, cache result = conjure.invoke( "levenshtein", s1="kitten", s2="sitting" ) # Returns: 3 # Next call: 0.3ms cache hit
Why
No dependencies, no trust chain
Zero imports
Generated code uses only Python builtins. No import statements, no eval, no exec. Enforced by AST analysis before caching.
Verified
Every example in the spec must pass in a sandbox. Fails? Error feedback goes back to the model. Three attempts before rejection.
Cached
Content-addressed via SHA-256. Same spec, same result. Sub-millisecond after the first call. 170,000x faster than generation.
ConjureEval-100
Benchmarked across 20 categories
| Application | Dependencies | With Conjure | Reduction |
|---|---|---|---|
| Flask blog | 13 transitive | 0 | 15x |
| FastAPI service | 15 transitive | 0 | 17x |
| CLI data tool | 5 transitive | 0 | 6x |
| Web scraper | 17 transitive | 0 | 20x |
| File sync utility | 8 transitive | 0 | 9x |
Average: 13x LOC reduction
Runtime
Runs entirely on your machine
No API keys. No cloud. No network requests. The model runs on-device via MLX on Apple Silicon.
Model
Qwen3.5-9B
OptiQ 4-bit quantization
Memory
5 GB
Mixed-precision weights
Cache speed
0.3 ms
170,000x vs generation