Zero-dependency programming

Conjure code
from specifications

Replace library imports with LLM-generated, verified, import-free code. One model binary. Zero transitive dependencies. Zero CVEs.

$ pip install conjure-llm click to copy
PyPI Paper (PDF)
The Problem
Every import is attack surface

A single compromised package in your dependency tree can affect your entire application. Conjure eliminates the dependency tree entirely.

Traditional

200 packages you didn't write

App
→ pip install
→ 200 packages
→ 1,500 transitive deps
→ 8M LOC of stranger code
→ dozens of maintainers
→ any one compromised = you're compromised
Conjure

One model, zero dependencies

App
→ YAML spec files
→ embedded LLM
→ verified code
→ cached forever
→ trust chain: 1 model binary
→ 0 CVEs, 0 maintainers to trust
The SQLite of libraries — embed the compiler, eliminate the supply chain
Pipeline
How it works
1

Write a YAML spec

Describe the function you need — signature, examples, constraints. Human-readable, version-controlled, auditable.

yaml spec: edit_distance function: levenshtein input: s1: str s2: str output: int examples: - input: { s1: "kitten", s2: "sitting" } output: 3 constraints: no_imports: true
2

Conjure generates & verifies

An embedded 9B model generates a self-contained implementation. AST analysis enforces zero imports. All examples must pass in a sandboxed environment.

python import conjure # First call: generate → verify → cache (~10s) result = conjure.invoke("levenshtein", s1="kitten", s2="sitting") # Returns: 3
3

Cache hit — instant

Second call uses the content-addressed cache. Sub-millisecond execution, 170,000× faster than generation.

python # Second call: cache hit (0.3ms) result = conjure.invoke("levenshtein", s1="hello", s2="world") # 170,000× speedup over generation
ConjureEval-100
87.9% correct on 100 specs

Evaluated across 20 categories — encoding, parsing, algorithms, math, collections, and more. Results improve from 30 to 100 specs.

pass@1
70.0%
70 / 100 — first attempt
pass@3
87.9%
87 / 99 — best of 3 attempts
Application Dependencies With Conjure LOC Reduction
Flask blog 13 transitive 0 15×
FastAPI service 15 transitive 0 17×
CLI data tool 5 transitive 0 6×
Web scraper 17 transitive 0 20×
File sync utility 8 transitive 0 9×
Average: 13× LOC reduction across 5 real Python applications
Translation Feasibility
How much can Conjure replace?
52%

of a typical Flask application's imports can be automatically identified as replaceable by Conjure specifications — including json, hashlib, base64, csv, statistics, collections, and itertools.

Security
Enforced, not hoped for

Security properties are structural guarantees verified by AST analysis — not heuristics, not trust.

🔒

No imports — ever

AST walker rejects any import, from-import, or __import__. Syntactic guarantee, not a policy.

🚫

No eval / exec

Dynamic code execution blocked at the AST level. No code injection vectors.

🏗️

Sandboxed execution

Restricted builtins, thread-based timeout. Generated code can't access files, network, or OS.

🔑

Content-addressed cache

SHA-256(spec + model + seed). Change anything and the cache invalidates. Deterministic builds.

Model
Runs on your machine

No API keys. No cloud. No network. The model runs entirely on-device via MLX on Apple Silicon.

Model
Qwen3.5-9B
OptiQ 4-bit quantization
Memory
5 GB
Mixed-precision weights
Cache hit
0.3 ms
170,000× vs generation