SimpleLLMs is a suite of agentic behaviors designed to transform Claude Code from a chat interface into a production-grade autonomous engineering team.
Inspired by the original R.A.L.P.H. pattern, this suite introduces specialized logic loops for research, creative pivoting, system integration, security auditing, and massive-scale processing—each encoded with a specific Model DNA for maximum effectiveness.
While all SimpleLLMs are model-selectable, they are rooted in the "DNA" of their origin engines. This bimodal architecture ensures a perfectly balanced autonomous team.
Precision, Reasoning, & User Experience
- L.I.S.A., B.A.R.T., M.A.R.G.E.
- Strengths: Deep reasoning, nuanced refactoring, empathetic UI/UX, and complex decision-making.
Scale, Security, & Compliance
- H.O.M.E.R., H.O.U.N.D., M.A.G.G.I.E.
- Strengths: Massive context windows (1.5M+ tokens), high-throughput parallel processing, adversarial security benchmarks, and low-latency compliance monitoring.
Unlike standard "blind loops," SimpleLLMs agents are grounded.
- Grounded Synthesis: Integrate with NotebookLM MCP to distill project documentation, PDFs, and whitepapers into a "Source of Truth."
- Specialized Execution: Select the agent behavior that matches your current bottleneck (e.g., Use L.I.S.A. for research-heavy features or B.A.R.T. for creative debugging).
These agents form the primary operational unit, leveraging specialized personas for specific development phases.
| Agent | DNA Engine | Role | Repository | Best For |
|---|---|---|---|---|
| L.I.S.A. | Claude | Research | lisa-agent | Deep-diving into complex docs and NotebookLM insights |
| B.A.R.T. | Claude | Innovation | bart-agent | High-entropy creative pivots and breaking through blocks |
| M.A.R.G.E. | Claude | Integration | marge-agent | Large-scale refactoring and system orchestration |
| H.O.M.E.R. | Gemini | Scale | homer-agent | Brute-force parallel processing and massive context handling |
| M.A.G.G.I.E. | Gemini | Compliance | maggie-agent | Human-in-the-loop oversight and SOP enforcement |
Discreet, purpose-built tools for security, quality assurance, and persistent memory.
| Agent/Tool | DNA Engine | Role | Repository | Best For |
|---|---|---|---|---|
| H.O.U.N.D. | Gemini | Security | hound-agent | Adversarial testing and red-teaming exploits |
| C.L.A.U.D.O.G. | Claude | QA | claudog | Gamified bug hunting and rigorous web-layer testing |
| R.A.L.P.H. | Claude | Persistence | ralph-agent | Relentless Automated Loop Processing Heuristic - "keep trying until it passes" |
| Blackboard | Shared | Governance | simplellms-blackboard | Registry of anti-patterns and cross-agent memory |
| Wiki | Shared | Context | claude-code-wiki | Project indexing and high-level architectural memory |
# Clone and install any agent
git clone https://github.com/midnightnow/lisa-agent.git
cd lisa-agent && ./install.shFollow the NotebookLM MCP Setup Guide to ground your agents in your documentation.
# Research-first development
simplellms --lisa "Implement the new authentication service"
# Epic orchestrator (M.A.R.G.E.) — plan, dispatch in parallel worktrees, review, pivot
simplellms --marge plan "Build an auth-gated dashboard that lists and creates agents"
simplellms --marge execute --concurrency 3
# Creative-pivot one-shot (still works; also invoked internally by M.A.R.G.E. on stuck tickets)
simplellms --bart "Find a way around this dependency conflict"
# Batch processing
simplellms --homer "Refactor all components to TypeScript strict mode"
# Security audit
simplellms --hound "Scan this repo for vulnerabilities"
# Check compliance
simplellms --blackboard "Verify all agents are behaving"┌─────────────────────────────────────────────────────────────────┐
│ SIMPLELLMS DECISION TREE │
├─────────────────────────────────────────────────────────────────┤
│ │
│ "What's blocking you?" │
│ │
│ ├── Need to understand before coding? │
│ │ └── L.I.S.A. → Research first, then implement │
│ │ │
│ ├── Stuck on the same error repeatedly? │
│ │ └── B.A.R.T. → Creative pivots and alternative paths │
│ │ │
│ ├── Multiple systems fighting each other? │
│ │ └── M.A.R.G.E. → Reconcile and guard execution │
│ │ │
│ ├── Need to process massive codebase fast? │
│ │ └── H.O.M.E.R. → Parallel batch operations │
│ │ │
│ ├── Concerned about UI or UX regressions? (Claude-powered) │
│ │ └── C.L.A.U.D.O.G. → Playwright-driven bug hunting │
│ │ │
│ ├── Worried about security or exploits? (Gemini-powered) │
│ │ └── H.O.U.N.D. → Adversarial testing & exploits │
│ │ │
│ ├── Need compliance oversight or "SOP" enforcement? │
│ │ └── M.A.G.G.I.E. → Human-in-the-loop & Safety Guardian │
│ │ │
│ ├── Lost in the codebase context? │
│ │ └── Wiki → Self-healing indexing & search │
│ │ │
│ └── Simple task, just need persistence? │
│ └── R.A.L.P.H. → Loop until it works │
│ │
└─────────────────────────────────────────────────────────────────┘
M.A.R.G.E.'s scope grows from "reconcile fighting systems" to epic orchestrator for the whole suite. She drives the other agents as pipeline stages and runs end-to-end without babysitting:
epic → L.I.S.A.-plan → H.O.M.E.R.-dispatch (parallel worktrees)
→ M.A.G.G.I.E.-review → B.A.R.T.-pivot (if stuck) → repeat
Task-Master-compatible verbs:
simplellms --marge init # in any git repo
simplellms --marge plan "<epic>" # or: --marge parse ./EPIC.md
simplellms --marge list # ticket graph w/ status
simplellms --marge execute --concurrency 3 # the whole loop
simplellms --marge show T001 # ticket details
simplellms --marge summary # progress dashboardState lives in .marge/ inside the target repo (plain JSON + markdown, git-friendly). Pluggable backends: MARGE_AGENT=claude|gemini|amp|echo. See src/marge/README.md for the full pipeline, schema, and env vars; see EPIC.md for the v0.2 roadmap expressed as a Marge spec sheet.
Design note. The classical Blackboard Pattern separates shared state (the board) from the controller that schedules knowledge sources. That mirror here is deliberate: .marge/epic.json is the board; M.A.R.G.E. is the controller; L.I.S.A. / H.O.M.E.R. / M.A.G.G.I.E. / B.A.R.T. are the knowledge sources. The existing SimpleLLMs Blackboard stays in its passive-governance role.
B.A.R.T. stays B.A.R.T. — the creative-pivot one-shot (simplellms --bart "<task>") still works, and M.A.R.G.E. calls the same persona internally when a reviewer returns needs_pivot.
SimpleLLMs implements and extends the autonomous loop pattern pioneered by:
- Geoffrey Huntley's Ralph Concept
- snarktank/ralph (Amp CLI implementation)
- NotebookLM MCP (Research grounding)
ghuntley.com/ralph (Concept)
│
▼
snarktank/ralph (Amp CLI Implementation)
│
▼
SimpleLLMs (Claude Code Extension)
├── R.A.L.P.H. ← Direct port and extension of Ralph pattern (fork of [snarktank/ralph](https://github.com/snarktank/ralph))
├── B.A.R.T. ← + Creative pivot strategy
├── L.I.S.A. ← + Research-first + NotebookLM
├── M.A.R.G.E. ← + Integration/cleanup focus
├── H.O.U.N.D. ← + Adversarial security testing (Gemini Engine)
├── C.L.A.U.D.O.G. ← + Playwright bug hunting (Claude Engine)
├── M.A.G.G.I.E. ← + Compliance & Safety Guardian (Gemini Engine)
├── Blackboard ← + Anti-pattern & Governance registry
└── Wiki ← + Agentic project indexing & context engine
MIT - Use freely. Build faster. Loop with purpose.
SimpleLLMs uses a federated Blackboard system to track and enforce agent behaviors.
Every violation recorded on the board follows this structure to ensure multi-agent compatibility:
### I WILL NOT [CONCISE_SUMMARY]
- **Severity**: [Critical | High | Medium | Low]
- **Agent**: [Agent Name | Universal]
- **Context**: [Description of the failing scenario]
- **Root Cause**: [Why the agent took this path (e.g., hallucination, missing check)]
- **Corrected Behavior**: [What the agent MUST do instead]
- **Evidence**: [Link to failing log or code snippet]SimpleLLMs - Simple LLM Suite Nine specialized agents, one powerful ecosystem.