Give your app exactly the data it needs — before your backend exists.

Describe whatever endpoint your app in natural language, drop in an OpenAPI spec — Zero-1 spins up live REST endpoints with realistic seeded data, full CRUD, and FK relationships. Tracks denormalized copies, propagates schema changes across existing data, and warns before a change would break a reference. Pair it with Claude Code to become an execution layer — all AI reasoning happens in your session, Zero-1 handles data, routing, and state locally. Export the schema when you're ready for your agent to build a working backend.

❯ I'm building a marketplace, I need sellers, products under /sellers/:id/products,
  and reviews under /products/:id/reviews. seed it heavy,
  like 5000 products and 10k reviews. gate writes with Bearer auth.

3 resources · 15100 records

· sellers    100   records
· products   5000  records  path: /sellers/:sellerId/products
· reviews    10000 records  path: /products/:productId/reviews

↳ products.sellerId  → sellers.id
↳ reviews.productId  → products.id

header guard: POST|PUT|PATCH|DELETE * — Authorization: Bearer required

Install

From source:

git clone https://github.com/thomscoder/zero-1 && cd zero-1

ANTHROPIC_API_KEY=sk-ant-... bun run cli.ts
# → http://localhost:7777  +  z1 › REPL

Or install the CLI:

curl -fsSL https://raw.githubusercontent.com/thomscoder/zero-1/main/scripts/install.sh | sh

Three modes

1. Standalone

Run Zero-1 as an HTTP server. Describe scenarios in the REPL or via API. One AI call per intent — schema parsing only. All data generation, CRUD, filtering, and graph tracking run locally with zero API calls.

z1                    # start + REPL
z1 --model gpt-4o     # use a specific model
ANTHROPIC_API_KEY=sk-ant-... bun run cli.ts

2. Claude Code native — z1 launch claude

Zero-1's primary mode. Claude Code IS the AI. Zero-1 executes.

z1 launch claude
z1 launch claude --model claude-opus-4-6

Starts the Zero-1 server and spawns Claude Code with the MCP pre-wired. Claude Code handles all reasoning and schema decisions. Zero-1 receives structured plans and executes them locally — zero redundant AI calls. Every token in your Claude Code session does actual work.

3. Scaffold export

Export your Zero-1 schema as a working production backend.

z1 export --scaffold hono --out ./my-app   # generate from .z1/state.json
z1 patch hono                              # sync changes (last-write-wins)

Or from the REPL:

z1 › .export --scaffold hono
z1 › .patch hono

Generates a complete Hono project: typed routes, in-memory data seeded with real records, TypeScript interfaces, and a /sync slash command for Claude Code. When your real backend is ready, the scaffold is the spec.

Features

Provider presets

Load a real provider's API shape instantly — no spec hunting, no key required:

# List available presets
GET /__generate/import/preset

# Load Stripe (customers, products, prices, payment intents, subscriptions at /v1/* paths)
POST /__generate/import/preset
{ "provider": "stripe" }

# Load GitHub (users, repos, issues at /api/v3/* paths + OAuth pre-configured)
POST /__generate/import/preset
{ "provider": "github" }

Available: github · google · stripe · linkedin · shopify

Each preset loads the correct real-world endpoint paths, realistic field shapes, and a pre-configured OAuth flow for that provider.

OpenAPI import

Point Zero-1 at any OpenAPI 3.x or Swagger 2.0 spec — URL or raw JSON:

POST /__generate/import/openapi
{ "url": "https://petstore3.swagger.io/api/v3/openapi.json" }
# or
{ "spec": { ...raw spec object... } }

Handles $ref resolution (recursive), allOf merges, oneOf/anyOf variants, and any path depth. Data is generated locally from the schema — no AI call.

REST

One description (or import) registers all resources with full CRUD — GET, POST, PUT, PATCH, DELETE. IDs are auto-incremented. FK references are always valid (data generated in dependency order).

Filter, search, sort — all query params, stackable:

GET /posts?role=admin&q=launch&sort=likes:desc&limit=10

Custom path patterns

Resources can live at any URL shape:

/api/v1/users
/payments/:paymentId/invoices
/orgs/:orgId/teams/:teamId/members

Named segments (:param) become FK equality filters on GET collection and are auto-populated on POST. The item ID is always the trailing segment: GET /payments/42/invoices/7.

Reactive triggers

"...when an order is created, subtract quantity from product stock. when stock hits 0, flip status to out_of_stock."

Describe side-effect logic inline. Supported operations: increment/decrement a field, set a value. Triggers fire synchronously on every matching mutation.

Schema evolution

"...add a views counter, rename likes to reactions, drop published. require X-API-Key on write requests."

Zero-1 sends only the schema to the AI (never the records), applies the diff, and backfills existing data locally. Added fields get context-aware generated values. Renamed fields preserve their values.

Pagination

"...cursor-based, 20 per page, use after as cursor param, edges for data array"

Configure the envelope your real API will produce. Zero-1 matches it. Supports page, offset, and cursor styles with customisable param names and response shapes.

Auth guards & response headers

"...require X-API-Key on all write requests to posts. inject X-Request-Id into every response."

Per-resource request guards and response header injection. Your frontend gets real 401s and real rejection bodies — no auth server needed.

OAuth 2.0 / OIDC

"...authorization code flow with PKCE. scopes: read, write, admin. JWT tokens, 1 hour expiry."

Full mock OAuth 2.0 / OIDC provider. Registers authorization, token, revoke, introspect, userinfo, and discovery endpoints. Per-client scope enforcement. Token signing via Web Crypto API (HS256) — no JWT library.

Chaos mode

"...30% error rate on POST /orders and 500ms–2s latency on all requests"

{ "resource": "orders", "method": "POST", "errorRate": 0.3, "errorCode": 503 }
{ "resource": "*", "method": "*", "latency": { "min": 200, "max": 800 } }
{ "resource": "users", "rateLimit": { "requests": 10, "windowSeconds": 60 } }

Knowledge graph

Zero-1 maintains a live graph of every resource — FK edges, denormalized copies, observed field types.

Write-through propagation — every PATCH/PUT walks the provenance graph and updates denormalized copies automatically.

Impact analysis — before applying a schema change, ask what breaks:

"What breaks if I switch the plan id to a string slug?"

Returns which FK references would stop resolving and which denormalized copies would go stale. Via MCP this is part of the natural workflow; via the API it's POST /__generate/graph/analyze.

2FA / OTP

Add two-factor authentication to any resource inline:

"...users with 2FA on login. OTP expires in 5 minutes."

Zero-1 registers POST /users/login and POST /users/verify endpoints. OTP codes are cryptographically random, 6-digit, single-use. Inspect pending codes at GET /__generate/2fa/pending.

JWT-based data scoping

"...each user should only see their own posts"

Per-resource per-user data isolation using JWT claims. Collection GETs filter automatically, item GETs return 403 for wrong owner, POSTs auto-set the owner field from the token.

MCP server

Zero-1 exposes its full surface as MCP tools for Claude Code and other AI agents.

Manual setup (if not using z1 launch claude):

// .claude/mcp.json
{
  "mcpServers": {
    "zero-1": {
      "command": "bun",
      "args": ["run", "/path/to/zero-1/mcp.ts", "--url", "http://localhost:7777"]
    }
  }
}

Tools include: create_scenario, evolve_schema, get_schema, list_resources, inspect_graph, analyze_impact, configure_headers, configure_oauth, configure_pagination, add_chaos, snapshot, reset, export_openapi, get_2fa_codes, and more.

Zero-waste architecture: every tool that normally triggers an AI call accepts a structured bypass. When called from Claude Code, Zero-1 skips its own AI calls entirely — Claude Code produces the plan, Zero-1 executes it. One AI session. No overhead.

AI providers

Zero-1 is bring-your-own-key. One AI call per intent — schema parsing only. Everything else is local.

# Anthropic (default)
ANTHROPIC_API_KEY=sk-ant-... bun run cli.ts

# OpenAI or any compatible endpoint
AI_PROVIDER=openai OPENAI_API_KEY=sk-... bun run cli.ts

# Ollama (local, no key needed)
AI_PROVIDER=ollama bun run cli.ts

Configuration

CLI flags:

z1 [--key sk-ant-...] [--port 7777] [--model claude-opus-4-6] [--web] ["inline prompt"]
z1 launch claude [--port 7777] [--model claude-opus-4-6]

API reference

Resource management

Dynamic CRUD (per resource)

Custom paths (e.g. /payments/:paymentId/invoices) work identically — named segments filter by FK equality.

OAuth 2.0

Knowledge graph

Infrastructure

Web console

http://localhost:7777/__generate — create scenarios, browse and edit records, test OAuth flows, manage chaos rules, view the live request log, export/import state.

Tests

Requirements

• Bun runtime
• An AI provider key (ANTHROPIC_API_KEY, OPENAI_API_KEY, or Ollama running locally)
• For z1 launch claude: Claude Code installed