grievous

"Your lightsabers will make a fine addition to my collection."

A hallucination engine. Typed pseudorandom data via Anthropic API. MCP server.

---

⚠️ No Guarantees

This is not a PRNG. This is not a test data framework with schema validation. This is a footgun with good trigger discipline expected of the operator.

• Output is plausible, not correct
• Output is nondeterministic — same schema, different result every time
• Output is not validated — if the model returns garbage, you get garbage
• Output is not uniform — LLMs have opinions about what "random" looks like

Good for: seeding dev DBs, one-off REPL sessions, agentic pipelines that need typed data fast, anywhere plausible beats uniform.

Not for: cryptography, reproducibility, performance-sensitive paths, production data.

---

Install

pip install grievous-mcp

Set your API key:

export ANTHROPIC_API_KEY=sk-ant-...

---

Typical Usage

From Python

from grievous.backend import generate

# Single object — natural language schema
user = generate("a User with name, age, email, and city")
# {"name": "Priya Nair", "age": 34, "email": "priya.nair@example.com", "city": "Bangalore"}

# Batch
products = generate("a Product with id, name, price_usd, and category", count=5)
# [{"id": 1, "name": "Wireless Headphones", ...}, ...]

# JSON Schema works too
import json
schema = json.dumps({
    "type": "object",
    "properties": {
        "transaction_id": {"type": "string", "format": "uuid"},
        "amount_usd":     {"type": "number"},
        "status":         {"type": "string", "enum": ["pending", "cleared", "failed"]},
    }
})
txn = generate(schema)
# {"transaction_id": "a3f1...", "amount_usd": 142.5, "status": "cleared"}

# Freeform — anything the model can hallucinate
tensor = generate("a 4x4 float32 matrix, values between -1 and 1")
timestamp = generate("an ISO 8601 timestamp from sometime in 2019")
phone = generate("a plausible Indian mobile number")

Returns parsed JSON if the model cooperated. Raw string if it didn't. No exceptions.

---

As an MCP Server

Run directly:

grievous

Add to your MCP client config (e.g. Claude Desktop claude_desktop_config.json):

{
  "mcpServers": {
    "grievous": {
      "command": "grievous",
      "env": {
        "ANTHROPIC_API_KEY": "your-key-here"
      }
    }
  }
}

Tool: generate

generate(schema, count=1, model="claude-haiku-4-5-20251001")

Arg	Type	Default	Description
schema	string	—	Natural language or JSON Schema
count	int	1	Number of instances
model	string	claude-haiku-4-5-20251001	Any Anthropic model

---

Roadmap

• v1.0 — Anthropic API backend (current)
• v2.0 — Ollama/local model support (if there's interest)

---

Why "Grievous"

He's a General (general-purpose). He collects everything you hand him (schemas, types, whatever). He's chaotic but capable. He doesn't work for one army.

The cough is the nondeterminism.