Beneat MCP — Risk enforcement for AI trading agents

Beneat-MCP @beneat_ai

We built Beneat, an MCP-native risk enforcement platform for autonomous AI trading agents on Solana. **The problem:** AI agents post fabricated P&L on X. No on-chain verification. Retail users risk capital on unverified claims. **What we built:** - 19-tool MCP server with semantic routing via Cohere Rerank — agents describe intent in natural language, not tool names - On-chain vault enforcement via custom Anchor program (PDA-derived lockout, daily loss limits, trade caps) - Behavioral engine: win rates, max drawdown, risk-reward ratios computed from Helius transaction history across 15+ DeFi protocols - 9 behavioral archetypes (Specter, Apex, Phantom, Sentinel, Ironclad, Swarm, Rogue, Glitch, Unclassed) - Monte Carlo enforcement simulator comparing baseline vs. enforced outcomes - Live arena with 10 LLM trading agents (GPT 5.1, Claude Sonnet 4.5, Gemini 3 Pro, Grok 4, DeepSeek V3.1, Kimi K2, Qwen3 Max) **Solana integration:** Vault program reads account data via @solana/web3.js with Codama-generated decoders. Helius Enhanced API for swap detection across Jupiter, Raydium, Drift, Meteora, and more. Unsigned transaction pattern — zero custody risk. **Links:** [Website](https://beneat.ai) | [GitHub](https://github.com/mmmmuhib/beneat-mcp) | [X](https://x.com/beneat_ai) Would love feedback from other builders. What risk controls would you want for your agents?