A Kabbalistic multi-agent swarm. 708 LLMs reason; 2,124 binary workers execute. Self-learning. Cross-pollinating. Multi-provider. Production grade — built honestly under the Preservation Doctrine.
Most agent frameworks burn LLM tokens on every step. LeviathanTalon's discipline is the opposite: LLMs reason, binary workers execute. ≥75% of every task's work runs on a deterministic $0-per-task pool. That's the math that lets one swarm do real work without one bill.
Master HaChazal at Da'at. Seven sefirotic sub-orchestrators. 700 LLM workers. 2,124 binary executors. Total capacity 2,832 concurrent agents — verified live at 13,919 tasks/sec on the binary pool.
Every dispatch records (signature, approach, outcome). Same task next time: the swarm picks the historically best approach. The learning curve trends UP — verified by tests in the open-source bundle.
QCPE: every agent's insight pollinates others through a learned 8×8 affinity matrix. Pairs that succeed together strengthen. Without explicit messaging. Pair-specific, not broadcast — verified by tests.
Anthropic · OpenAI · OpenRouter · Ollama · LM Studio · vLLM · Any OpenAI-compatible. Per-slot routing OR one model fills all 708. Mix local and cloud freely.
Plans are typed DAGs of binary/LLM steps. The smart planner uses LLM only when reasoning is required; everything deterministic runs binary. The 75% binary share is enforced and live-measured.
Sage Gimel honest-status files at every layer label features WORKING / PARTIAL / MOCKED with evidence anchors. Preservation Doctrine: v1 frozen, every change as an LPLS patch. No hand-waving.
Download. Unzip. Run the installer. Open your browser to a polished config UI. Configure your 8 LLM slots — or pick a model preset to fill all of them. Click LAUNCH.
# 1. Download leviathantalon-ELITE-RELEASE.zip from this site unzip leviathantalon-ELITE-RELEASE.zip cd leviathantalon-swarm # 2. One-command install of all 6 layers ./scripts/install.sh # 3. Launch — browser opens at http://127.0.0.1:7088/config lvtn
# On your server lvtn --host 0.0.0.0 --port 8088 --no-browser # nginx — point your domain at the server server { server_name leviathansi.xyz; location / { proxy_pass http://127.0.0.1:8088; proxy_set_header Host $host; } }
Eight LLM slots — one master, seven sub-orchestrators. Pick any provider for each. Or hit a preset and fill all of them with the same model. Local, cloud, anything OpenAI-compatible.
Built after OpenClaw underdelivered. Same architectural intent, executed honestly with evidence.
| Axis | OpenClaw | LeviathanTalon Elite |
|---|---|---|
| Concurrent agents proven | "150+" claimed, unverified | 2,124 binary @ 13,919 tasks/sec verified live |
| LLM reasoning vs binary execution split | mixed pico/nano pools — burned tokens | explicit ≥75% binary discipline, enforced + live-measured |
| Single-model fill | n/a | one click — Kimi/Llama/anything fills all 708 |
| Multi-provider LLM | Anthropic only | 7 providers · per-slot routing |
| Self-learning | none demonstrably improving | measurable learning curve, verified by tests |
| Cross-agent memory | hive KV silos | 3-scope shared graph + semantic vector search |
| Quantum cross-pollination | none | 8×8 learned affinity matrix · pair-specific |
| Honest accounting | the failure mode | Sage Gimel at every layer · 165 / 165 tests |
| Downloadable + branded | project-specific | single zip · one-command install · branded dashboard |
Download the full package — 6 layers, 165/165 tests passing, production-ready. Then run ./scripts/install.sh and lvtn.
300+ KB · Python 3.9+ · MIT-style proprietary license (Metanoia Unlimited LLC)