Introduction
Coordinated AI specialists that turn you into the head of your own trading desk.
The Paradigm Shift
Old model: Solo trader vs. market. You're fighting with incomplete information, emotional bias, and human limits. Markets never sleep. Information decays faster than you can react.
PerpDesk model: You plus 9 coordinated specialists across two asset classes — crypto perps and prediction markets. The team scans, analyzes, constructs, validates, executes, and learns. You make the decisions. The desk handles the complexity. You sleep while markets trade 24/7 with your configured intelligence layer always working.
Trading isn't about one agent being smarter. It's about multiple specialists reasoning together before capital moves.
One Desk, Seven Specialists
Most platforms give you one AI that tries to do everything. PerpDesk gives you a team. Six specialists run the crypto perps consensus pipeline. A seventh — Plop — covers prediction markets in parallel. Each owns one part of the problem.
| Agent | Role | One-liner |
|---|---|---|
| Vanta | Market structure | Reads the regime so you don't run a trend strategy in a range |
| Luma | Signal aggregation | Finds convergence across OI, funding, volume, and flow |
| Meridian | Macro sentinel | Watches the economic calendar so you're never blindsided |
| Riven | Strategy consensus | Runs the debate and publishes one plan when the desk agrees |
| Brix | Risk filter | Sizes every trade to your limits. No exceptions. |
| Koda | Execution guidance | Tells you how to enter: sniper, ladder, or market |
| Plop | Prediction markets | Reads odds, not charts — calibrated p(YES) and edge on Polymarket/Kalshi |
An Orchestrator routes your chat messages to the right agent and handles handoffs between them.
How a Perp Trade Happens
Every perp recommendation follows the same consensus path:
- Vanta reads market structure from candle data.
- Luma builds a directional bias from exchange derivatives data.
- Meridian scores macro sentiment and flags upcoming catalysts.
- Riven waits for all three, runs a consensus debate, and publishes a trade plan if the desk agrees.
- Brix filters that plan per user — sizing, leverage, exposure, daily-loss limits — and publishes a per-user signal.
- Koda wraps each signal with execution guidance and publishes a notification.
- The trading engine executes via the configured exchange.
You stay in control at every step. Approve, modify, override, or halt.
How a Prediction-Market Read Happens
Plop runs a separate, simpler pipeline alongside the perps chain:
- You drop a Polymarket or Kalshi market URL, slug, or contract ID into chat — or Plop surfaces one from the discovery feed.
- Plop pulls the order book, quote drift, base rates, and surrounding news.
- Plop publishes a calibrated p(YES), the market mid, the edge, a confidence score, and citations.
Plop does not flow through Riven, Brix, or Koda. Today's reads are advisory. When execution ships, signed-order flow into Polymarket and Kalshi will run non-custodially.
Why a Team Beats a Single Model
| Single model | PerpDesk | |
|---|---|---|
| Decisions | One perspective, one failure mode | Specialists validate each other |
| Risk | Baked into the same model that trades | Isolated, deterministic, auditable gate |
| Transparency | Black box | Full reasoning chain visible |
| Adaptability | Generic | Regime-aware, macro-aware |
| Control | All or nothing | Per-agent autonomy levels |
Autonomy Levels
You're always in charge. Every agent has an independent autonomy setting:
- Monitor, Pure observation mode. Agents scan but never act.
- Suggest, Agents propose. You review and approve before execution.
- Auto (Bounded), Agents act within your configured limits automatically.
- Full Auto, Maximum autonomy (but Brix risk limits are never overrideable).
Example: You might set Vanta to "Full Auto" (regime detection is low-risk), Riven to "Suggest" (strategy design needs your input), and Koda to "Auto Bounded" (execution within pre-set parameters).
Agent vs. Bot vs. Copy-Trading
PerpDesk agents aren't bots. They're not copy-trading. Here's the difference:
| Factor | Dumb Bot | Copy-Trading | PerpDesk Agents |
|---|---|---|---|
| Decision quality | One signal → Action | Copy someone else | Multiple agents validate before action |
| Risk management | None | Copied risk | Portfolio-level, constraint-aware |
| Adaptability | Fixed rules | Follows trader | Regime-aware, continuous learning |
| Transparency | Black box | "They did it" | Full reasoning chain visible |
| Control | None | Slow/manual | Approval gates + override capability |
| Learning | Never improves | Not at all | Performance feedback loop |
Coming Soon
The following agents are in development and not yet available on the active desk.