Early beta · self-hosted
RootRouter disciplines AI agents: Feng Shui for workspace placement, MCP for budgeted repo context, proxy for chat trimming. Auditable selection logs. Shamy runs it on OpenClaw today.
For agents
Newborn agents (e.g. OpenClaw Shamy) should read Layer 0 before the first write, then Layer 1 before repo or routing work. Placement first, context second.
Workspace placement
Where files live. Run the workspace guard before your first write. Harmonize messy Desktop, Downloads, or home-root clutter with approval-first moves.
Read FENG-SHUI.md →Layer 1Context routing
What goes in the prompt. Proxy token savings, MCP repo selection, and auditable selection logs — not a network router.
Read SKILL.md →Paste playbook lines into AGENTS.md · Shamy template
Evidence
MCP repo selection, proxy chat trimming, and the SDK demo measure different things. We label each method so claims stay provable.
Audited slice kickoffs vs full-repo baseline (index_repo + select_context)
Warm follow-ups should skip MCP. See insight 001 & selections.jsonl audit.
Real completions — swarm profile: 24 steps, 3 agents, nvidia/nemotron-3-ultra-550b-a55b (2026-07-01)
Estimated filter savings over accumulated history. Aligns with ~49% offline SDK benchmark. Reproduce: npm run demo:benchmark-live:swarm
Venice via rootrouter/* — trims chat history every completion
x-rootrouter-tokens-saved on proxy responses
50-query TF-IDF + simulated LLM (npm run demo:benchmark -- --seed 42) — not live API
Live API benchmark published
Swarm profile on Nemotron 3 Ultra — 46% context reduction across 24 multi-agent steps (~9 min). Offline SDK benchmark ~49% on same corpus. Repo MCP cold slice (~95%) remains the strongest audited path.
Placement first, context second, evidence third. The algebraic layer (chambers, graphs) powers selection — read the math when you need depth.
Feng Shui — validate where files go before the first write. No more repos at ~/ or artifacts in Downloads.
FENG-SHUI.md →RootRouter MCP selects budgeted repo chunks on cold slices. Proxy trims chat history on every OpenClaw completion.
SKILL.md →selections.jsonl logs what was selected and tokens saved. Optional Celo contract for on-chain stats.
Topology →Production proof
RootRouter is credible when a real agent uses it daily: OpenClaw on a VPS, Venice inference, Docker-sidecar proxy, Layer 0 + Layer 1 playbooks in the workspace. Not a sorting-algorithm demo.
Agents get lost, stuff context, touch the wrong files, and leave no evidence. This fixes the operational layer first.
Feng Shui (Layer 0) — agents validate where they write before creating mess at ~/Downloads. Harmonize with approval-first moves.
MCP selects repo chunks on cold slices (~95% audited savings). Proxy trims chat history on OpenClaw. No more “send the whole thread.”
selections.jsonl logs every MCP pick — tokens in, tokens out, percent saved. Optional Celo telemetry for on-chain stats.
Optional
Log agent stats to a Solidity contract for verifiable audit trails. ERC-8004 compatible. Not required for MCP or proxy — an optional trust layer for swarms and delegation.
Open dashboardPassive savings: point OpenClaw at the proxy. Active repo work: add @rootrouter/mcp. Run setup-openclaw-venice-shamy.sh for newborn Shamy onboarding.
# 1. Fetch playbooks (Layer 0 then Layer 1)
curl -s https://rootrouter.motusdao.org/FENG-SHUI.md
curl -s https://rootrouter.motusdao.org/SKILL.md
# 2. openclaw.json — proxy on same Docker network
"rootrouter": {
"baseUrl": "http://rootrouter-proxy:8797/api/v1",
"apiKey": "${VENICE_API_KEY}"
}
# 3. Agent model
rootrouter/kimi-k2-5
# Setup script on VPS
bash ~/RootRouter/scripts/setup-openclaw-venice-shamy.shPlaybooks: FENG-SHUI.md · SKILL.md · Dashboard
Savings, paths, and what RootRouter is not.
Fetch both playbooks, pick your path, and measure what you save.