Spec-driven backend platform for FastAPI
Generate REST APIs, GraphQL, search, version history, and an admin UI from a single Python model.

Renamed from autocrud to specstar (v0.10.0). The old name still installs as a deprecation shim that redirects every autocrud[.X] import to specstar[.X]. New projects should pip install specstar. See the migration guide.

v0.11 adds spec-driven authoring (additive, zero breaking changes). Edit a single spec.md in prose; a Claude Code skill or the specstar gen CLI translates it into the declarative Python the engine consumes. See the Spec-Driven Authoring guide and the v0.11 upgrade notes.

Focus on business logic, not infrastructure.

⭐ If you find this project useful, consider giving it a star.

Why SpecStar

Modern backend development repeatedly rebuilds the same infrastructure:

• CRUD APIs
• validation
• search and filtering
• version history
• permissions
• background jobs
• admin tools

Most of this code is not business logic.

SpecStar eliminates this repetition by using a spec-driven architecture.

Define your model once, and the framework generates the rest.

What sets SpecStar apart

Most "model → API" generators stop at routes. SpecStar treats every resource as a versioned timeline by default: updates create revisions, history stays queryable, and rollback is built in. Don't need history? Ignore it — the same routes work as plain CRUD.

Example

Define a resource model:

from msgspec import Struct

class User(Struct):
    name: str
    email: str

Register the model:

from fastapi import FastAPI
from specstar import spec

app = FastAPI()

spec.add_model(User)
spec.apply(app)

spec.add_model(User) is shorthand for spec.add_model(Schema(User)) — the canonical form once you start tracking schema versions or migrations is spec.add_model(Schema(User, "v1")), which the quickstart docs use. Both run; pick whichever fits your stage.

Start the server:

uvicorn main:app

Optional startup tuning:

export SPECSTAR_DEFAULT_QUERY_LIMIT=1000

This controls the default page size for list endpoints (GET /{model} and search routes). The built-in fallback is 2**32 - 1 (effectively unlimited) — pick a sane value for production. Per-request limit and offset query params still override it.

You now automatically get:

POST   /user
GET    /user
GET    /user/{resource_id}
PUT    /user/{resource_id}
PATCH  /user/{resource_id}
DELETE /user/{resource_id}

The path segment follows model_naming (default: kebab-case of the model name, e.g. User -> /user). PUT replaces the whole resource; PATCH expects a JSON Patch (RFC 6902) array of operations rather than a partial object.

OpenAPI documentation is generated automatically.

Architecture

graph TD

FastAPI --> SpecStar
SpecStar --> ResourceManager
ResourceManager --> Storage

Storage --> MetaStore
Storage --> RevisionStore
Storage --> BlobStore

SpecStar --> REST_API
SpecStar --> GraphQL_API
SpecStar --> UI_Generator

Core Features

Spec-driven APIs

SpecStar generates APIs directly from Python models.

Model
  ↓
REST API
GraphQL API   (opt-in: pip install specstar[graphql] + add_route_template(GraphQLRouteTemplate()))
OpenAPI

REST and OpenAPI are wired up automatically when you call spec.apply(app). GraphQL is opt-in: install the extra (pip install specstar[graphql]) and register the template explicitly:

from specstar import spec
from specstar.crud.route_templates.graphql import GraphQLRouteTemplate

spec.add_route_template(GraphQLRouteTemplate())

Versioned resources

Every resource maintains immutable revision history.

Resource
 ├── r1
 ├── r2
 └── r3

Advantages:

• audit history
• rollback
• draft workflows
• debugging

Built-in search

Search operates on indexed metadata instead of scanning full resource payloads.

QueryBuilder
   ↓
ResourceManager.search()
   ↓
MetaStore.search()

Background jobs

Jobs are modeled as resources.

create()
  ↓
message_queue.put(resource_id)

Workers process jobs through:

ResourceManager.start_consume()

Storage abstraction

SpecStar supports multiple storage backends.

The rows above are typical profiles, not the only valid combinations. The three storage layers (IMetaStore / IResourceStore / IBlobStore) are independent — for example, a Postgres resource_store exists (resource_data table with data BYTEA), so you can keep revision payloads in Postgres instead of S3 if that fits your deployment. You can also implement custom storage systems.

UI generation

SpecStar can generate a web interface directly from the API.

API
 ↓
UI generator
 ↓
admin dashboard

This allows rapid creation of internal tools.

Comparison

Quickstart

Install:

pip install specstar

Spec-driven (v0.11+) — bootstrap a starter project, then describe resources in prose:

specstar init my_app
# edit spec.md, then in Claude Code:
/specstar regen
# CI-friendly drift check (no LLM):
specstar verify

See the Spec-Driven Authoring guide for the full workflow.

Classic Python — register models directly:

Run your app:

uvicorn main:app

Open:

http://localhost:8000/docs

Documentation

Full documentation:

https://hychou0515.github.io/specstar/

Example use cases

SpecStar works well for:

• internal tools
• content systems
• configuration management
• job processing systems
• administrative APIs
• workflow management systems

License

MIT