walid@portfolio:~/lab/agent-trading-desk$
cd../lab
02ideaSep 2026

Nine agents, one voice

A voice-driven trading desk where a master agent routes your spoken command to nine specialists

Most people cloning Jarvis build something general. This one goes narrow and deep into a single domain: you speak, a master agent routes the work to nine specialists — macro, smart money, recon, analyst, technician, quant, risk, execution, journal — and the charts on the wall obey out loud. The architectural idea worth stealing is the gate: analysis cannot execute, and execution cannot authorise itself.

AgentsTradingVoiceClaude CodeElectronTradingViewTradingAgentsworldmonitorhermes-agentwhisper.cppKokoro TTSFinnhubMediaPipe
i
Narrow beats general

Everyone is cloning a general-purpose Jarvis — an assistant that does a little of everything and nothing particularly well. The alternative is to go deep in exactly one world. Here that world is the markets: you give one spoken command, a master agent decides which specialists should answer, and the answer comes back in their voices while the charts flip and the levels get drawn. It is a research and awareness co-pilot, not an autotrader.

The HUD

What it looks like running

Nine agent seats around the orb, a live signal log down the left, the session clock and agent load down the right. Worth handing to Claude Code as a visual reference — it reads images, so the layout, glow and panel structure can be matched from the screenshot alone.

Mid-session, orb in focus mode. Market pulse, the NYSE session and the signal log are live; the calendar dots mark FOMC and options expiry.
The cast

One master, nine specialists

The master never does specialist work itself — it runs the desk. That separation is the whole design.

01·THE MASTER

Orchestrator and risk router

Takes one command, routes the work to the right agents, synthesises what comes back, and answers out loud. It is the only agent that talks to you: it delegates down and summarises up.

Tier 1

Intelligence — what is happening

AgentBeatWhat it owns
AtlasMacroReads the Fed, rates and the market’s weather — the tide everything else floats on.
CapitolSmart moneyTracks what Congress, insiders and the big funds are quietly buying, before the news does.
ScoutReconFinds whatever stock is going viral today — ahead of the feed.
Tier 2

Analysis — what it is worth

AgentBeatWhat it owns
AthenaAnalystReads the filings and calls the structure on the chart — supply, demand, liquidity.
ChartistTechnicianOwns the screens. Flips symbols and marks the lines: where to buy, and where to run.
OracleQuantRuns the forecasting models and answers in probabilities, never promises.
Tier 3

Execution — where money moves

AgentBeatWhat it owns
SentinelRisk officerSizes every position, sets every stop, holds veto power. Its favourite word is no.
PilotExecutionPlaces the trade — but only once Sentinel has signed off.
LedgerThe bookJournals every call the desk makes, especially the bad ones, and watches system health.
The idea

Four rules make it a desk instead of nine chatbots

01·WRITTEN BY AGENT

Describe an agent, get the service

The whole thing is built and driven with Claude Code — you describe an agent in plain English and it writes the service. That is what turns “an AI that talks” into a team that does work.

02·ONE TO MANY

Each specialist is its own service

Single job, clean interface to the others, and only the master talks to you. Adding a tenth agent later does not touch the other nine.

03·DRIVES THE SCREEN

The charts are part of the desk

The technician does not describe the chart, it drives it — Claude connects to TradingView through an open-source MCP server, so “pull up Nvidia and mark the levels” is a command rather than a request.

04·THE GATE

Exactly one path to money

A proposal only becomes a trade after the risk officer stamps it. Analysis cannot execute; execution cannot self-authorise. The risk officer is built into the architecture rather than bolted on — this is the rule worth copying even if you never build the other eight agents.

!
What it is, and what it is not

A research and awareness co-pilot with a built-in risk officer — not a money printer. It surfaces structure and context; it does not tell you what happens next. Educational, and not financial advice. The orchestrator also lives or dies on reasoning quality: routing one spoken command to the right specialists, and knowing when to stop and let the risk officer say no, is the part a weaker model gets wrong.

The shopping list

Every piece of it is free

The stack that looks like a film set is, piece by piece, software you can install tonight.

01·THE SHELL

An Electron app — a desktop web page

A dark HUD, nine agent seats, one glowing orb. The agent writes all of it; the visual taste is the only part you supply.

02·THE VOICES · $0

Local text-to-speech

The open-source Kokoro engine gives every agent a distinct voice — dozens of them, roughly two seconds a clip, entirely on your machine. A paid service is the polish upgrade, not the starting requirement.

03·THE EARS · $0

Local speech recognition

whisper.cpp runs speech-to-text on your own hardware. The app hears you finish a sentence and acts — no cloud, no per-minute fees.

04·THE DATA · $0

Live quotes and history

A free Finnhub key covers quotes, movers and news; yfinance covers history. Every number the desk speaks should come from a real feed — a desk that invents prices is a toy.

05·THE SCREENS · $0

Chart takeover

The open-source TradingView MCP server lets the desk flip symbols and draw on your real charts.

06·THE HANDS · $0

Optional gesture steering

MediaPipe hand tracking runs in the browser, which is what lets you steer the orb without touching anything.

Prompt 1

The starter prompt

Open an empty folder and start here, then iterate one agent at a time.

starter-prompt.txt9 lines
Build me an Electron app called "my trading desk":
- a dark cinematic HUD with 9 agent seats arranged around a glowing central orb
- each agent has a name, a role (macro / smart-money / recon / analyst /
  technician / quant / risk / execution / journal) and its own TTS voice
- a master agent that takes my typed or spoken command, decides which
  specialist should answer, and speaks the reply out loud
- live quotes from the Finnhub API (I'll paste my free key)
- one hard rule: any "trade" action must be approved by the risk agent first
Start with the HUD and two agents; we'll add the rest one by one.
The panels

What is real on that screen

In the original build every market panel is wired to a live feed rather than placeholder text, and the alerts come from its own sub-agents.

PanelWired to
Market pulse + session chipsLive quotes from the free data tier, refreshed every minute, red when down.
Session clockThe real NYSE phase and countdown, computed from New York time.
Signal logAlerts streamed from the sub-agents — macro, recon, risk.
Agent rosterNine seats here; tailor the nodes to your own agents. Three is plenty to start.
Prompt 2

Rebuilding the HUD from the screenshot

Attach the image above and paste this. The last line matters most: build the static layout first, then wire one panel at a time.

hud-prompt.txt10 lines
Here's a screenshot of a trading desk HUD (attached).
Use it as the visual reference for my desk's HUD:
- same layout: left column (market pulse, task queue, signal log),
  a glowing central orb with my agent nodes around it,
  right column (market session, calendar, agent load)
- wire the market panels to REAL data: Finnhub quotes for the pulse
  rings, the live NYSE session + countdown, my event calendar
- the signal log streams alerts from MY sub-agents: [list your agents]
- never show invented numbers on a market panel — if a feed is down, show "—"
Build the static HUD first, then wire one panel at a time.
The open-source stack

Three repos carry the heavy lifting

Star counts checked against GitHub on 15 Sep 2026 — all three have grown since the write-up quoted them.

RepoStarsWhat it gives you
TradingAgents106kAn analyst team that argues: bull vs. bear debate → trader → risk manager → verdict.
worldmonitor86kThe open-source war room — live news, conflicts, shipping, the whole world on one screen. The macro agent runs on it.
hermes-agent246kThe learning brain that orchestrates it all, and does not lose context along the way.
The hard part

What actually turns nine agents into a desk

The roster is the easy half. These are the pieces that decide whether it feels alive or feels like nine chatbots talking over each other.

Orchestration that routes one command to the right agents and merges their answers into a single reply.The chart takeover — flipping symbols and drawing levels on a real chart while you are still talking.A voice pipeline that gives each agent its own personality and, just as importantly, its own timing.Zero-lag delivery, so the desk answers the instant you stop speaking rather than a beat later.Hand tracking, if you want to steer the orb without a keyboard.Every agent wired to real, verified market data — never invented numbers.
i
Straight talk

This is not a download. There is no installer and no repo to clone — the nine agents run on custom tooling, skills and backend wiring that would not copy-paste into anyone else’s setup. What is genuinely reusable is the design: the orchestrator-and-specialists split, the single gated path to execution, and the free building blocks above. Build your own shape of it — three agents might be plenty.

← previous
The agent that makes the video
next →
The agentic backend
← all experiments