archiet-microcodegen 0.2.2

Archiet's core algorithm in one file: PRD text -> FastAPI app ZIP, pure stdlib, zero dependencies.

archiet-microcodegen

PRD text → working FastAPI + PostgreSQL app → ZIP. One command. No LLM. No API key. Pure Python stdlib. 1,377 lines you can read in 20 minutes.

Inspired by Karpathy's micrograd: this file is the complete algorithm. Everything else is just efficiency on top.

pip install archiet-microcodegen
archiet-microcodegen prd.md --out ./myapp/
cd myapp && docker compose up
# → http://localhost:8000/docs

What you get

Write a PRD describing your entities and user stories. Get back a bootable app:

File	What it does
main.py	FastAPI app with /docs OpenAPI explorer
models/*.py	SQLAlchemy 2.0 models, one per entity
schemas/*.py	Pydantic v2 Create / Update / Response schemas
routers/*.py	Full CRUD — list, create, get, update, delete
auth.py	JWT via httpOnly cookies — register / login / logout / me
alembic/	Migrations pre-configured, run alembic upgrade head
tests/test_*.py	pytest suite with DB override, auto-skip if Postgres unreachable
docker-compose.yml	Postgres 16 with healthcheck-gated startup
ARCHITECTURE.md	ArchiMate 3.2 element map — ApplicationComponent, DataObject, ApplicationService, ApplicationInterface
openapi.yaml	Machine-readable API contract, importable into Postman

Every entity has per-tenant data isolation built in. Every query filters by the authenticated user's ID. No data leaks between users.

Zero LLM calls. Zero API keys. Zero runtime dependencies. Pure Python stdlib.

The four stages

PRD text
  │
  ▼ parse_prd()        — regex extraction: entities, fields, user stories, integrations
manifest dict
  │
  ▼ manifest_to_genome() — maps to canonical IR with ArchiMate 3.2 element typing
genome dict              (same schema as archiet.com's full platform)
  │
  ▼ render_genome()    — string.Template FastAPI rendering
{path: content}
  │
  ▼ pack()             — stdlib zipfile
ZIP bytes

The genome is the key: your PRD becomes an ArchiMate 3.2 architecture document before any code is generated. That's what makes the output traceable and maintainable — not just scaffolded.

As a library

from archiet_microcodegen import microcodegen, parse_prd, manifest_to_genome, render_genome

# Full pipeline
zip_bytes = microcodegen(open("prd.md").read())

# Or stage by stage
manifest = parse_prd(prd_text)
genome   = manifest_to_genome(manifest)
files    = render_genome(genome)   # dict[str, str]
# inspect files["ARCHITECTURE.md"] to see the ArchiMate element map

Example PRD

# Task Manager

## Entities
- Task: title (string, required), description (text), status (string), due_date (date)
- Project: name (string, required), description (text)

## User Stories
As a user, I want to create tasks so I can track my work.
As a user, I want to assign tasks to projects so I can organise them.

## Integrations
- Stripe for billing

Run it: archiet-microcodegen prd.md --out ./taskapp/ — you'll get 23 files including a fully wired FastAPI app with Task and Project CRUD, auth, migrations, tests, and an ARCHITECTURE.md with ArchiMate 3.2 notation.

Why no LLMs

LLMs are great at understanding messy PRDs. They're unnecessary for the generation step — once you have a clean manifest, code emission is deterministic. This file handles the deterministic half. The full platform at archiet.com handles the LLM-powered extraction, 14 stacks, frontend, mobile, and delivery gates.

If a bug doesn't reproduce with microcodegen, it's in an efficiency layer, not the algorithm.

What's NOT here

• No LLM extraction (full platform handles messy natural-language PRDs)
• No frontend, no mobile, no Stripe wiring, no rate limiting, no audit logging
• No multi-stack (FastAPI only here — for Flask, Django, NestJS, Go, Java see archiet.com)

License

MIT. See LICENSE.

Links

• Source: github.com/aniekanasuquookono-web/archiet
• Issues: github.com/aniekanasuquookono-web/archiet/issues
• Full platform: archiet.com