Build a trading swarm on Robinhood Chain with three agents

A trading swarm is a few AI agents that split one trade into separate jobs. This one has three. A Scout finds Robinhood Chain tokens that are moving. A Skeptic checks each one and can veto it. A Trader paper-trades whatever survives and writes every entry and exit to a file.

The whole thing is a free template for Claude Code. It reads its data from the PumpPill MCP server, which is free and needs no account. It trades on paper only. There is no wallet, no private key and no order code anywhere in it.

This guide walks through how it works, how to run it, and how to change its rules.

Why three agents

More agents don't create an edge. They make the checks stricter.

A single agent that finds a token and then approves it is grading its own homework. It already likes the token by the time it checks it. Splitting the work fixes that. The agent that finds a token is never the agent that approves it, and the approving agent starts from "no".

That is the only thing the swarm adds. It won't find a signal a single agent couldn't. It makes it harder for a bad token to slip through on a good story.

Step 1: Get the template and connect it

You need Claude Code installed. Clone the template, open the folder and start Claude Code:

git clone https://github.com/JackedBased/pumppill-swarm-template

cd pumppill-swarm-template

claude

The folder has a .mcp.json file that points at the PumpPill MCP server. Claude Code asks you to approve it the first time. Type /mcp to check that pumppill shows as connected.

That's the whole setup. There are no keys to paste and nothing to install.

Step 2: Meet the Scout

The Scout's job is to find tokens that are moving now, with real buyers, after their launch.

It skips anything younger than 60 minutes on purpose. PumpPill's own research on Robinhood Chain found that momentum after launch was the signal that held up. Signals from the first minutes of a launch did not. So the Scout never picks a token because it is new.

It reads two lists from the MCP. The first is rh_ignition_calls, PumpPill's momentum calls on breakouts with real buyers. The second is rh_new_launches over the last 24 hours, keeping only tokens whose price is up over the last hour.

Then it confirms each candidate with rh_token_snapshot. A candidate stays only if its pool is deep enough, at least 20 different wallets bought it in the last day, and more wallets bought than sold. Official Robinhood stock tokens are left out.

The Scout returns at most five candidates, each with one sentence on why. An empty list is a normal answer.

Step 3: Meet the Skeptic

The Skeptic's job is to find the reason not to buy. It runs five checks in order and stops at the first failure.

1. Pre-trade check. It calls pre_trade_check, which returns one word: clear, caution, avoid or unknown. Avoid is a veto. Unknown is also a veto, because no stored read means no evidence.
2. Holders. The token needs at least 1,000 holders.
3. Top-10 share. The 10 largest wallets can hold at most 35% of supply. If the Skeptic can't find the number, it vetoes.
4. Holder spread. Buyers in the last day must outnumber sellers, with at least 20 buyers.
5. Exit. It calls exit_check. If the token can't be sold, or the deployer can still mint or freeze it, that's a veto.

A veto is final for the round. Nothing in the loop can overrule it, including the orchestrator.

Step 4: Meet the Trader

The Trader has no wallet and no tool that can place an order. The project's settings file denies the MCP's trading tools outright, so it can't call them even by mistake.

For each approved token, it reads the market cap from rh_token_snapshot and writes an entry line to paper/trades.jsonl with a fixed paper size.

Every round, it marks open positions first. A position closes on the first of four things: the take-profit, the stop, the time limit, or the token stopping trading. On Robinhood Chain, a pulled pool doesn't show up as a trade. The tape just stops. So a token with no trades in 24 hours is recorded as a total loss.

The market cap is a mark. Real fills are worse. The Trader subtracts a cost assumption from every result so the paper numbers don't flatter you.

Step 5: Run a round

Type /swarm in Claude Code. One round runs through four gates.

1. Target. The round states what it's looking for before any tool call: how many candidates, how many open positions, which settings.
2. Observe. The Trader marks open positions, the Scout finds candidates, the Skeptic vets them, and the Trader opens paper positions for the approved ones. You see each agent's full output before the next one starts.
3. Bounded repair. If an agent returns broken output, it gets sent back once with the exact problem. If it fails again, that step is dropped for the round. There's never a third try.
4. Stop. The round ends when the Trader is done, the Scout found nothing, the Skeptic vetoed everything, or a repair failed twice. You get a five-line report.

/swarm 3 runs three rounds in a row. The command caps a single run at 10.

Step 6: Change the rules

Every threshold lives in one table in CLAUDE.md: minimum market cap, liquidity, holders, the top-10 cap, paper size, take-profit, stop, time limit and the cost assumption. Edit the table and the agents use the new numbers on the next round.

The agents are plain-English files in .claude/agents/. You can add a check to the Skeptic by writing one more numbered step. Keep each check to something a tool actually returns. A check the agent has to guess at is worse than no check.

Run it on paper for a while before you trust any rule. Read the misses as closely as the wins.

Free, and what a data key adds

Everything above runs on the free MCP. Two parts get better with a PumpPill data API key set as PUMPPILL_DATA_KEY.

• The Scout gets live momentum calls. On the free MCP, Ignition calls arrive after a delay. The data API's signal feed delivers them live on the Builder and Pro plans.
• The Skeptic gets the top-10 share as a number. Without a key, it reads the top-10 line from the safety read's text and vetoes when it can't find one. With a key, it reads top10_pct straight from the token endpoint.

The template works without a key. A key means fewer vetoes for missing data and fresher candidates.

Where to go next

If you'd rather not run code, the PumpPill trade bot on Telegram has Agent Squads. A squad buys only when 2 of 3 of your agents agree, and it starts on paper.

Template: github.com/JackedBased/pumppill-swarm-template

Data API and keys: pumppill.org/data

Trade bot: t.me/PumpPill_Trade_Bot