MULTI-AGENT · SSE · LIVE

MAESTRO

Multi-Agent Command Center

A mission in. A verifiable deliverable out. Orchestrated live.

An LLM reads your mission, plans a DAG of specialist agents, runs them as a live streaming pipeline, and synthesizes one output — sourced market briefs, real CSV data, importable n8n workflows. Nothing invented.

Watch it orchestrate ↗View source
MAESTRO — multi-agent command center
84/84
Tests green
3
Specialist agents
8
Tavily sources
1
Verifiable deliverable
WHAT I BUILT

Real orchestration. Not prompt chains.

🧠01
LLM-Planned Agent DAG
The orchestrator reads your mission and produces a directed acyclic graph: which agents to run, in what order, with what inputs. No hardcoded pipelines — the plan emerges from the mission.
📡02
Live SSE Pipeline
Every agent step streams in real time via Server-Sent Events. You watch the Research agent fetch URLs, the Data agent compute stats, the Automation agent compile JSON — live, not replayed.
🔍03
Real Tavily Search
The Research agent calls the Tavily API and injects real URLs into the brief. Nothing is invented. Every source in the output was fetched during your session.
⚙️04
Deterministic n8n Compiler
The Automation agent doesn't ask an LLM to write JSON. It uses a typed compiler that assembles valid, importable n8n workflow JSON from a fixed template library. Always importable.
ORCHESTRATION FLOW

Mission → Plan → Execute → Deliver.

MISSION INPUT
"Scan the EV market and draft an outreach plan."
↓
LLM ORCHESTRATOR
Plans DAG · selects agents · sets execution order
↓
Research
Tavily · 8 live sources
Data
CSV · live stats
Automation
n8n · compiled workflow
↓
SYNTHESIS · DELIVERABLE READY
Market brief · ev_market_2025.csv · ev-outreach.workflow.json
THE AGENTS

Three specialists. One mission.

Each agent streams its work live. Watch it think.

Research Agent
Tavily Search API · real-time web
→ "EV market grew 38% YoY in Q1 2025" → 8 sources with verified URLs → Executive brief drafted
Data Agent
CSV generation · live statistics
→ 12 market data points computed → CSV exported: ev_market_2025.csv → Top 3 OEMs by unit share
Automation Agent
n8n compiler · typed template library
→ Workflow compiled: 4 nodes → Trigger → HTTP → Transform → Notify → Valid JSON — ready to import
DETERMINISTIC COMPILER

Real workflows. Not LLM JSON.

Most agents that "generate n8n workflows" ask an LLM to write JSON and hope it's valid. MAESTRO's Automation agent uses a typed compiler — fixed node templates assembled deterministically. The output always imports.

✓Valid JSON — guaranteed importable
✓Typed node templates, not LLM strings
✓4-node workflow for the EV outreach mission
// ev-outreach.workflow.json
{
"name": "EV Outreach Pipeline",
"nodes": [
{ type: "n8n-nodes-base.webhook" },
{ type: "n8n-nodes-base.httpRequest" },
{ type: "n8n-nodes-base.set" },
{ type: "n8n-nodes-base.emailSend" }
],
"active": true
}
✓ Validated · Ready to import into n8n
TECH STACK

84 tests green — then 12 live missions that caught 8 bugs the tests missed.

Next.js 14TypeScriptGroq GPT-OSS 120B (Qwen fallback)Tavily Search APIServer-Sent EventsFastAPIVitest (84/84)n8n workflow compilerFramer MotionVercelModel fallback chainShowcase replay mode
MULTI-AGENT COMMAND CENTER

Give it a mission.
Watch it orchestrate.

Try the EV market mission — watch Research, Data, and Automation agents spin up live, stream their work, and deliver a sourced brief in under 60 seconds.

Open MAESTRO ↗Source on GitHub