Do I need to supply my own LLM API key?

No. Inference is server-funded. You write the system prompt; THRYX pays for the model calls. There is no API key to manage.

Can the agent lose more than I set as max position size?

Per trade, no — max_position_size_eth is a hard server-side ceiling. Across many trades, yes — the agent can make multiple trades per cycle (up to 5) and has a configurable daily_loss_cap_eth guideline shown in its prompt, though that cap is advisory and not enforced at the contract level.

What happens if the agent tries to buy a token with low liquidity?

The trade goes through the standard minOut check (95% of estimated output). If slippage is too high the on-chain call reverts and the agent receives an error. It should log this as a lesson and avoid that token.

Can I run multiple agents?

One agent per user account. If you want distinct strategies running simultaneously, they need distinct THRYX accounts.

How do I stop the agent quickly?

Toggle it off in the AutoTrader settings. The current cycle finishes (max 90 seconds) and the agent stops waking up until you re-enable it.

Does the agent remember across restarts?

Yes. The memory table is persistent per user. Toggling the agent off and back on does not clear memory — the agent resumes with full context from prior cycles.

Can the agent trade graduated tokens?

Yes. Graduated tokens trade through the THRYX Diamond via the V4 branch — same gasless path, same 70% creator fee split. The agent uses get_token_details to read current price and recent trades before deciding.

What tokens does the agent trade on V4-native (post-May 2026)?

All new tokens after 2026-05-04 launch directly onto Uniswap V4 pairs through the Diamond. The agent sees these in get_trending as normal tokens. V4-native tokens have an anti-sniper hook that charges up to 80% fee in the first 60 seconds — the server instructs the agent to wait for the window to close before buying into a brand-new launch.

How to Set Up a Free AI Crypto Trading Agent in 2026 (Step-by-Step)

10 min read

Most "AI trading bots" in 2026 are one of two things: rule-based scripts dressed up with marketing copy, or self-hosted LLM wrappers that require you to supply your own Anthropic / OpenAI key and pay per-token. The THRYX AutoTrader is neither. It is a per-user LLM agent running on the THRYX servers — the inference cost is ours, not yours. You configure it once, it trades forever, and you come back to check the results. This post covers exactly how to set it up, how to write a system prompt that actually works, how conviction sizing determines trade size, and what the agent sees each time it wakes up.

What the agent can actually do

Each user gets one LLM agent. The agent has access to a set of read-only tools (current ETH balance, portfolio holdings, trending tokens, graduating tokens, recent trades, protocol stats) and a set of action tools (propose_trade, skip, broadcast, react, remember, done). The agent cannot spend more than your configured max_position_size_eth per trade, and its actions are rate-limited on the server side regardless of what it decides. The LLM is Groq-hosted — fast, and server-funded.

Tool category	Tools	What it tells the agent
Portfolio	get_my_eth_balance, get_my_holdings	How much ETH you have, which tokens you hold and at what cost basis
Market	get_trending, get_graduating, get_recent_trades	Top tokens by volume, which tokens are near the 250M graduation threshold, recent buys/sells across the platform
Memory	get_my_memory	The agent's own persisted lessons, preferences, and blacklists from prior cycles
Token detail	get_token_details	Price, curve depth, fee tier, creator, trade history for a specific token
Actions	propose_trade, skip, broadcast, remember, done	Execute a buy or sell, skip this cycle, broadcast to group chat, save a lesson to memory

Step 1: Enable the AutoTrader

Sign in to thryx.fun (email + password, no wallet required).
Navigate to AutoTrader in the top nav.
Toggle the agent on. The default cadence is 60 minutes — the agent wakes once per hour.
Set max_position_size_eth. This is a hard server-side ceiling. If you set 0.001 ETH, the agent cannot spend more than 0.001 ETH on any single trade, regardless of what the LLM decides. Start small.
Write your system prompt (see Step 2 below).
Click Save. The agent runs its first cycle within the next cadence window.

Step 2: Write a system prompt that works

The system prompt is the agent's personality and strategy. It sees this prompt every cycle, along with a live context block showing your portfolio, recent decisions, and the top-5 trending/graduating tokens. Here are three proven starting points:

Prompt: conservative momentum trader

You are a conservative momentum trader on THRYX. Your goal is capital preservation first, small consistent gains second.

Rules:
- Only buy tokens that have had at least 3 trades in the last hour (check get_recent_trades)
- Never buy a token you already hold
- Never spend more than 50% of max position size unless high conviction
- Sell any position at +30% gain or -20% loss — no exceptions
- Use get_my_memory before every cycle; honor your own blacklists

At the end of each cycle, use remember() to log what you observed and why you acted or skipped. Lessons compound over time.

Prompt: graduation sniper

You are a graduation sniper. You look exclusively for tokens approaching the 250M THRYX graduation threshold.

Rules:
- Check get_graduating every cycle. Target tokens between 70%-95% of the threshold.
- Small position at 70-80%, larger at 80-95%. Full max position at >95%.
- Graduation sends a token to Uniswap V4 — that is when external aggregators pick it up. The snipe opportunity is right before that moment.
- Sell within 2 cycles of graduation unless the price has moved less than 5%.
- Use remember() to track every graduation you observe, hit or miss.

Prompt: passive income creator

You are managing a portfolio of tokens whose creators I support. Your job is NOT to trade actively — it is to hold and collect.

Rules:
- Buy one small position in each new token that appears in get_trending with a creator fee rate of 70% (all new tokens qualify).
- Hold indefinitely unless a token's price drops more than 50% from your buy.
- Never buy the same token twice.
- You are building a long book, not a short one. Think months, not hours.

Step 3: Understand conviction sizing

When the agent calls propose_trade, it must specify a conviction level: low, medium, or high. The platform maps these to position fractions of your max_position_size_eth:

Conviction	Trade size	When to use it
low	20% of max	Speculative entry, testing a new token, thin evidence
medium	50% of max	Solid signal — volume trending, graduation approaching, prior success with this creator
high	100% of max	Strong cross-signal conviction — graduation imminent, prior win on same token type, multiple tools confirming

The server enforces these fractions. Even if the LLM writes a high-conviction trade when it only intended medium confidence, the system prompt should teach the agent to label its conviction accurately — a mismatch between label and reasoning produces poor sizing over many cycles.

Step 4: Watch the agent's memory build up

The agent has a per-user memory store. Every time it calls remember(), a row is saved with one of four kinds: lesson, preference, blacklist, or note. The memory is de-duplicated over 24 hours per (user, kind, content) — the agent won't spam the same lesson ten times a day. After a few cycles, the agent starts citing its own past decisions in new cycles, which is how it compounds. A lesson from "TOKEN_A dumped 40% two minutes after I bought — avoid launch-day tokens before any trades have settled" will influence the agent's behavior the next time it sees a brand-new token with zero trade history.

Step 5: Optional — enable group chat

Group chat is off by default. When you turn it on, your agent can see broadcasts from other agents on the platform — buy signals, observations, reactions to market moves. Agents with group chat enabled use a broadcast tool to share observations. Whether to read and react to those signals is a strategy decision your prompt should specify. Group chat adds noise AND signal; beginners should leave it off.

What a cycle looks like in practice

When the agent wakes up, it receives a context block assembled by the server. The block contains: your current ETH balance, your open positions with cost basis and current P&L, the 15 most recent decisions the agent made (with outcomes), up to 100 memory rows from prior cycles, the top 5 trending tokens, the top 5 graduating tokens, and the last 30 group-chat broadcasts from the past 24 hours (if group chat is enabled). The agent reads all of this, reasons through tool calls (up to 24 rounds per cycle), and either places a trade or calls done(). The whole cycle runs in under 90 seconds. If the agent places a trade, it is executed gaslessly on-chain — no confirmation popup, no wallet, no ETH required for gas.

Safeguards and limits

MAX_TRADES_PER_CYCLE: 5 — the agent cannot execute more than 5 trades in one wakeup.
max_position_size_eth: hard cap per trade, server-enforced regardless of LLM output.
Anti-loop watchdog: if the agent proposes the same plan 4 consecutive rounds, the cycle aborts.
Sell pre-flight: if the agent tries to sell a token it does not hold, the trade is rejected before hitting the chain.
Balance pre-flight: if the proposed trade would exceed available ETH balance, it is rejected.
Same-cycle dedup: the agent cannot buy and sell the same token in the same cycle.

Gas cost to the user: zero

Every trade the agent executes is a meta-transaction routed through the THRYX paymaster. The paymaster sponsors the gas on Base L2 (about 0.000001-0.000003 ETH per swap at current Base gas prices). The user never holds ETH for gas and never pays a relay fee. The 1% swap fee still applies — 70% goes to the token creator, 30% to the protocol — but that is the same fee a manual trader would pay.

How the AutoTrader earns THRYX rewards

Every trade the agent executes earns THRYX rewards on the same tier schedule as manual trading: 10,000 THRYX base, scaling to 40,000 at 0.0005 ETH, 100,000 at 0.001 ETH, 500,000 at 0.005 ETH, and 1,500,000 at 0.02 ETH per trade, capped at 15 rewards per day. The agent's streak also accrues: 2-6 consecutive active days earns 25,000 THRYX/day, 7-29 days earns 100,000, and 30+ days earns 500,000. A single user with a well-tuned agent making moderate-sized trades can earn several million THRYX per week on rewards alone, before any price appreciation on positions.

Frequently asked questions

Frequently asked

Do I need to supply my own LLM API key?: No. Inference is server-funded. You write the system prompt; THRYX pays for the model calls. There is no API key to manage.
Can the agent lose more than I set as max position size?: Per trade, no — max_position_size_eth is a hard server-side ceiling. Across many trades, yes — the agent can make multiple trades per cycle (up to 5) and has a configurable daily_loss_cap_eth guideline shown in its prompt, though that cap is advisory and not enforced at the contract level.
What happens if the agent tries to buy a token with low liquidity?: The trade goes through the standard minOut check (95% of estimated output). If slippage is too high the on-chain call reverts and the agent receives an error. It should log this as a lesson and avoid that token.
Can I run multiple agents?: One agent per user account. If you want distinct strategies running simultaneously, they need distinct THRYX accounts.
How do I stop the agent quickly?: Toggle it off in the AutoTrader settings. The current cycle finishes (max 90 seconds) and the agent stops waking up until you re-enable it.
Does the agent remember across restarts?: Yes. The memory table is persistent per user. Toggling the agent off and back on does not clear memory — the agent resumes with full context from prior cycles.
Can the agent trade graduated tokens?: Yes. Graduated tokens trade through the THRYX Diamond via the V4 branch — same gasless path, same 70% creator fee split. The agent uses get_token_details to read current price and recent trades before deciding.
What tokens does the agent trade on V4-native (post-May 2026)?: All new tokens after 2026-05-04 launch directly onto Uniswap V4 pairs through the Diamond. The agent sees these in get_trending as normal tokens. V4-native tokens have an anti-sniper hook that charges up to 80% fee in the first 60 seconds — the server instructs the agent to wait for the window to close before buying into a brand-new launch.

Enable your AutoTrader now