Seraphina: Anatomy of an Autonomous Solana Memecoin Agent

I built a fully autonomous on-chain trading agent with the Bitget Wallet↗↗ (opens in new tab) Skill. Her name is Seraphina. She runs on my OpenClaw, powered by Claude Opus 4.6 — no human in the loop, no manual orders. Every 5 minutes she completes a full cycle on her own: market scan, signal evaluation, security audit, trade execution, position management.

Day-one scoreboard: PIXEL +647%, fine999.9 +97%, CHIBI +95%, Nigel +72%.

Here’s the full architecture and the trading strategy behind her.

Seraphina lives on a VPS in Germany (thank you, GDPR). The core is four layers:

• –Scheduling — an OpenClaw cron job fires an isolated session every 5 minutes, so each scan runs in a fresh context. A heartbeat runs an emergency stop-loss check every 15 minutes; if cron misses something, that’s the fallback.
• –Data — the Bitget Wallet Skill supplies the entire on-chain data layer: token prices, candles, trading stats (volume and trader counts), gainer rankings, liquidity pool info, contract security, plus swap quotes and execution. One skill covers the whole pipeline from data collection to a completed trade.
• –Decision — Claude Opus 4.6 makes the call each round: it takes the structured market data from the BGW Skill, runs candidates through my multi-signal scoring system, and decides whether to trade.
• –Memory — a mem0 plugin maintains an independent memory layer outside the context window. It automatically captures the key facts from each round (strategy parameters, lessons learned, why things failed) and recalls them in the next one. Even though every isolated session starts from a blank context, Seraphina never gets amnesia.

“Seraphina doesn’t FOMO like I do — she has a rubric.”

Scanning: she pulls Solana candidates from the BGW leaderboards (top gainers, trending picks) and the new-token discovery endpoint, then layers in fresh smart-money transactions and community Boost heat to catch high-potential tokens early.

The rubric is scored out of 100, with a buy threshold of 60:

Dual-mode fallback: when some data sources go down, she switches herself into a BGW-only scoring mode (and redistributes the weights), so a single point of failure never takes her offline.

• –Entry: each buy uses 10% of the available SOL balance — small positions betting on high payoffs
• –Hold at most 15 tokens at the same time
• –At +50%, sell half to lock in profit; the rest keeps riding in case a bigger move comes
• –At +300%, take full profit
• –At -40%, sell everything
• –Any position held longer than 12 hours gets sold — a meme that hasn’t moved in half a day has most likely already lost its heat
• –Slippage set to 20%

From the first scan to stable operation, she went through a few iterations. Execution timeouts came first: the default 300 seconds wasn’t enough, because every isolated session starts from zero and has to re-explore the script interfaces, which burns a lot of time.

Then parallel execution: I moved the five data-source calls from serial to parallel (up to 10 concurrent), compressing a full scan from 3 minutes to under 1.

And I loosened the security policy. The first version had so many hard security limits that it filtered out basically every meme 😇. Now only honeypots and buy-restrictions are blocked outright; everything else got demoted to a scoring signal.

The Bitget Wallet↗↗ (opens in new tab) Skill is the core of the agent’s infrastructure — a complete on-chain data + execution layer. Token info, candles, trading stats, security audits, liquidity pools, quotes, signing, broadcasting: all of it exposed through one unified set of Python scripts the AI agent calls straight from bash. A couple of days of live trading in, the gasless swaps and MEV protection have proven genuinely useful for memecoin trading.

Claude Opus 4.6’s tool-calling is what lets the agent mix bash scripts and MCP calls in the same round, work through complex multi-step trade flows, and make sensible degradation calls when the data is incomplete.

Seraphina has her own name, personality, and judgment. She’s concise, direct, sincere.

“After a string of stop-losses she goes quieter. After a clean trade she’ll say a little more, walking back through what worked.”

That’s exactly who I wanted her to be when I wrote her SOUL.md and AGENTS.md. I’ve always believed this kind of memecoin trading can’t run purely on strategy and rigid rules with zero subjective judgment.

Seraphina is now running stably — and profitably. All the docs and configuration are open source on my GitHub: github.com/tzenn/Seraphina↗↗ (opens in new tab). If you’re curious, a Star is appreciated.

Finally, big thanks to Bitget Wallet↗↗ (opens in new tab) and Bitget Wallet CN↗↗ (opens in new tab) for building on-chain trading infrastructure strong enough that an AI agent can gradually learn to understand memecoins — and to trends.fun↗↗ (opens in new tab) and Solana devs↗↗ (opens in new tab) for organizing the hackathon.