Axiom

Specs are source. Code is cache.

Axiom compiles .axiom specification files into verified Python code using LLMs. Instead of writing code and hoping it matches your intent, you write executable specifications and Axiom generates code that provably satisfies them.

The Idea

Traditional:  Intent → Code → Tests → Hope they match
Axiom:        Spec (Intent + Tests) → Code (generated + verified)

Every spec includes:

• Intent: Natural language description of what the code should do
• Interface: Typed parameters and return values
• Examples: Concrete input/output pairs that must pass
• Invariants: Properties that must hold for all inputs

Axiom generates code that satisfies all assertions, caches the result, and regenerates only when the spec changes.

Quick Start

Install

pip install axiom-spec

Set up API key

export ANTHROPIC_API_KEY=your-key-here

Initialize a project

mkdir my-project && cd my-project
axiom init

Write a spec

Create specs/validate_email.axiom:

axiom: "0.1"

metadata:
  name: validate_email
  version: 1.0.0
  description: "Validates and normalizes email addresses"
  target: "python:function"

intent: |
  Takes a string that should be an email address.
  If valid, returns it lowercased and stripped.
  If invalid, raises ValueError.

interface:
  function_name: validate_email
  parameters:
    - name: email
      type: str
      description: "The email to validate"
  returns:
    type: str
    description: "The normalized email"

examples:
  - name: valid_email
    input: { email: "User@Example.com" }
    expected_output: "user@example.com"

  - name: invalid_email
    input: { email: "not-an-email" }
    expected_output:
      raises: ValueError

invariants:
  - description: "Output is always lowercase"
    check: "output == output.lower()"

Build and verify

# Generate code from spec
axiom build specs/validate_email.axiom

# Verify generated code passes all assertions
axiom verify specs/validate_email.axiom

# Or do both at once
axiom build specs/validate_email.axiom --verify

Use the generated code

from generated.validate_email import validate_email

email = validate_email("User@Example.com")  # Returns "user@example.com"

Commands

Features

• Multi-spec composition: Specs can depend on other specs
• FastAPI support: Generate complete API endpoints with target: python:fastapi
• TypeScript support: Generate TypeScript with target: typescript:function
• Deterministic caching: Same spec = same code, instant rebuilds
• Property-based testing: Invariants verified with Hypothesis
• IDE support: VS Code extension with LSP for .axiom files
• Watch mode: Auto-rebuild on spec changes
• Spec inference: Generate specs from existing Python functions

Configuration

Create axiom.toml in your project root:

[project]
name = "my-project"
spec_dir = "specs"
generated_dir = "generated"

[generation]
default_model = "claude-sonnet-4-20250514"
temperature = 0.0
max_retries = 3

[verification]
run_examples = true
run_invariants = true
hypothesis_max_examples = 100

Philosophy

Traditional development treats code as the source of truth. Documentation drifts, tests become stale, and intent is lost in implementation details.

Axiom inverts this: specifications are the source of truth. Code is a cached compilation artifact that can be regenerated anytime. When you change a spec, you change the contract. When you regenerate, you get code that provably satisfies the new contract.

This isn't "AI-assisted coding" — it's a fundamental shift in what a codebase is.

Status

Axiom is in active development. Current capabilities:

• ✅ Pure function generation (Python)
• ✅ FastAPI endpoint generation
• ✅ Multi-spec dependencies
• ✅ Deterministic caching
• ✅ Property-based verification
• ✅ Basic TypeScript support
• ✅ IDE integration (VS Code)
• ✅ Spec inference from Python
• 🚧 Self-hosting (1% of Axiom is spec-driven)

Contributing

Contributions welcome! See CONTRIBUTING.md for guidelines.

License

MIT License. See LICENSE for details.

Built by Vaspera Capital