LLM-Powered Creative Idea Engine with Local Rule-Based Convergence
One call: `spark_ideate("pet supplies")` → your LLM diverges → 5-layer filter converges → structured output.
Works with: Claude Code · Codex CLI · Hermes Agent · OpenClaw
Star us on GitHub
This project has no relation to Apache Spark (big data engine) or GitHub Spark (low-code platform).
from agent_spark import spark_ideate
results = spark_ideate(
domain="pet supplies",
llm=lambda prompt, system, model: my_llm_call(prompt, system),
)
for r in results:
print(f"{r['title']} — {r['one_line']}")
print(f" Feasibility: {r['feasibility_score']}/5 Novelty: {r['novelty_score']}/5")One call, three steps:
- Load presets — 20 domains with pre-scanned pain points auto-detected
- LLM diverges — your model generates 6+ ideas from the domain + pain points
- Local filter converges — 5-layer rule engine drops hallucinated, buzzword-heavy, or market-duplicate ideas
No API key needed if you pass your own LLM. Just set OPENAI_API_KEY for auto HTTP mode.
| Other AI idea tools | Agent Spark |
|---|---|
| LLM prompt → raw text, you filter manually | 5-layer filter runs automatically after LLM |
| Pure LLM — no validation | L1 fact check, L4 market duplicate detection |
| Every idea passes through | Buzzword, vague, and hallucinated ideas dropped before you see them |
| One prompt for all | 20 domain presets enrich the LLM with real pain points |
The key insight: LLMs diverge (generate many ideas). Local rules converge (find the few that are real). Agent Spark does both.
Your LLM Agent Spark
──────────── ─────────────
| |
| generate_ideas() |
|───────────────────────────────▶│
| (system=diverge_prompt │
| user=domain+pain_points) │
│ │
│ JSON ideas returned │
│◀───────────────────────────────│
│ │
│ five_layer_filter()
│ L1: fact validation
│ L2: logic consistency
│ L3: technical feasibility
│ L4: market duplicate check
│ L5: buzzword / value check
│ │
│ Pass results returned │
│◀───────────────────────────────│
| Platform | Integration | How It Works |
|---|---|---|
| Claude Code | Auto-loads AGENTS.md |
from agent_spark import spark_ideate |
| Codex CLI | Auto-loads AGENTS.md |
Same import |
| Hermes Agent | integrations/hermes/SKILL.md |
Skill triggers full 6-round interview |
| OpenClaw | integrations/openclaw/ |
Plugin panel with wizard UI |
# From GitHub (zero external Python deps for filter engine)
pip install git+https://github.com/jzfjs2008-ship-it/agent-spark.git
# From PyPI (package name is offline-idea-spark)
pip install offline-idea-spark
# With optional REST API
pip install "offline-idea-spark[api]"Core engine: stdlib only. LLM client uses stdlib urllib. Optional REST API adds FastAPI + uvicorn + pydantic.
| Function | What It Does |
|---|---|
spark_ideate(domain, llm=...) |
Primary. LLM diverge + local converge in one call |
generate_ideas(domain, llm=...) |
Lower-level, more control |
find_domain("pet supplies") |
Look up 7 pre-scanned pain points |
Filter.run(ideas, ...) |
Run filter manually |
list_domains() |
All 20 preset domain names |
pip install -e ".[dev]"
pytest tests/ -v21 tests covering filter engine, LLM response parsing, preset integrity, and convenience API.
| Tool | Purpose |
|---|---|
| Crush Your Passion | Project risk assessment |
| KillAI-WinCleaner | AI environment cleanup |
Workflow: Agent Spark generates ideas → Crush Your Passion evaluates → KillAI-WinCleaner resets.
Add new domains to agent_spark/presets/domains.json. Every added domain enriches the LLM's pain-point context for all users.
See CONTRIBUTING.md and open issues via domain suggestion template.
Q: Is this the same as Apache Spark?
A: No. Apache Spark is a big data engine. This is an idea engine for AI coding agents.
Q: Does it work without an LLM?
A: Yes — find_domain("pet supplies") + Filter.run() works offline with pre-scanned pain points. But the full power is with an LLM.
Q: What does "zero dependencies" mean?
A: The core filter engine uses only Python stdlib. The LLM client also uses stdlib urllib (no OpenAI SDK needed).