Prob Desk

Kalshi multi-agent trading desk — Google ADK + Gemini with quant, risk, execution, and sentiment agents. Public market data out of the box; optional authenticated trading via the Kalshi SDK.

Architecture  ·  Design  ·  Web UI  ·  ADK docs

Google AI Studio key  ·  Kalshi demo setup (slides)

---

Quick start (web UI — recommended)

Requirements: Python 3.10+, Node 20+, a Google AI Studio API key.

git clone https://github.com/JasonMun7/ProbDesk.git
cd ProbDesk
./scripts/setup.sh

Edit .env at the repo root and set GOOGLE_API_KEY, then:

cd ui && npm run dev

Service	URL
Web desk	http://localhost:3000
Agent (AG-UI)	http://127.0.0.1:8000

The setup script installs pip install -e ".[ui]", copies .env.example → .env if needed, and runs npm install under ui/.

prob-desk web   # same as cd ui && npm run dev

(back to top)

---

Web desk features

Area	What you get
CopilotKit desk	Chat with the trading_director ADK agent; tool calls stream in the sidebar
Generative panels	Kalshi search, order book, portfolio, and trade receipts render in the center panel
HITL trading	approve_kalshi_order — Approve / Deny in chat before execution
Chat welcome	Starter suggestions and follow-ups tuned to desk workflows
Thread history	New chat, switch threads, delete — persisted in the browser
Get Started	Onboarding modal for first-time setup
Settings	Kalshi API, Google key, and AgentPhone status (reads repo-root .env + local desk settings)
Help	? in the desk header — shortcuts and demo prompts

(back to top)

---

Kalshi demo account

Public search and order books work against the demo API without keys. For portfolio, orders, and live execution, create demo API credentials and add them to .env.

Step	Action
1	Follow the Kalshi demo setup walkthrough (Google Slides)
2	Save your RSA private key as secrets/kalshi/private_key.pem (see secrets/kalshi/README.md)
3	Set KALSHI_API_KEY_ID, KALSHI_PRIVATE_KEY_PATH, and keep KALSHI_TRADE_API_BASE on the demo host
4	Restart the agent (cd ui && npm run dev) and confirm Settings → Kalshi API → Connected

(back to top)

---

AgentPhone (optional)

Voice / SMS via AgentPhone MCP tools on the trading director.

Config	Where
AGENTPHONE_API_KEY	.env at repo root
Agent ID (agt_…)	Settings → AgentPhone (saved under .prob-desk/desk-settings.json, gitignored)

Use the Kalshi demo setup slides for demo API keys before wiring AgentPhone flows in the desk.

(back to top)

---

Environment variables

Copy .env.example → .env at the repo root (the Next.js app loads this file automatically).

Variable	Required	Purpose
GOOGLE_API_KEY	Yes	Gemini via ADK
GOOGLE_GENAI_USE_VERTEXAI	—	Keep false for AI Studio
KALSHI_API_KEY_ID	For trading / portfolio	Authenticated Kalshi SDK
KALSHI_PRIVATE_KEY_PATH	For trading / portfolio	Path to RSA PEM (see secrets/kalshi/)
KALSHI_TRADE_API_BASE	—	Demo vs production API host
AGENTPHONE_API_KEY	No	AgentPhone MCP
AGENTPHONE_AGENT_ID	No	Optional in .env; prefer Settings UI
PROB_DESK_AGUI_URL	No	Override AG-UI URL (default http://127.0.0.1:8000/)

(back to top)

---

Run modes

Mode	Command	Use when
Web desk	cd ui && npm run dev or prob-desk web	Primary UI — chat, generative panels, trade approval
AG-UI only	prob-desk serve	Backend API for the UI (port 8000)
Terminal REPL	prob-desk	Quick prompts without the browser
ADK Web	prob-desk adk or ./scripts/run-adk-web.sh	Stock Google ADK dev UI (port 8501, not 8000)

Do not run ADK Web and the CopilotKit AG-UI server on the same port. Default layout: AG-UI on 8000, ADK Web on 8501.

More UI detail: ui/README.md.

(back to top)

---

Demo prompts (PGA example ticker)

Use market KXPGATOUR-PGC26-SSCH (Scottie Scheffler PGA) in chat:

1. Order book — Show the orderbook for KXPGATOUR-PGC26-SSCH
2. ~$1 yes buy — ask to spend about $1 on yes, use the book for price/size, approve_kalshi_order in chat, then kalshi_sdk_create_order after approval
3. Risk — For KXPGATOUR-PGC26-SSCH, check my portfolio and transfer_to_agent(agent_name='risk_manager') for a short risk summary

Trade flow: request in chat → Approve / Deny in the sidebar → executed trade receipt in the center panel.

(back to top)

---

Manual setup (without the script)

python3 -m venv .venv && source .venv/bin/activate
pip install -e ".[ui]"
cp .env.example .env
cd ui && npm install && npm run dev

Conda alternative: conda env create -f environment.yml && conda activate prob-desk.

(back to top)

---

Testing

pip install -e .
pip install pytest
pytest tests/ -v

(back to top)

---

Project layout

ProbDesk/
├── prob_desk/          # ADK agents, Kalshi tools, AG-UI server
├── ui/                 # Next.js + CopilotKit web desk
├── docs/readme-logo.png # README header logo
├── ui/public/brand/    # Web UI logo assets
├── scripts/setup.sh    # One-shot local setup
├── .env.example
└── DESIGN.md

Topic	Doc
Agent graph & tools	docs/adk-agents-architecture.md
Execution RL scripts	docs/tactical-policy-cli-scripts.md

(back to top)

---

License

MIT — see pyproject.toml.