Acc. II · Protocolum sylvestre — thirty-seven-server MCP grove, cultivated 2025
MCP Server Platform — DeFi tools for AI agents
De-facto lead engineer (took ownership after handover) · 2025 – 2026 · AltLayer
~37 protocol MCP servers (~176 tools across EVM + Solana) integrated into AltLLM's multi-model LLM platform, served through a unified stateless Hono gateway.
Protocolum sylvestre — thirty-seven-server MCP grove, cultivated 2025
Field notes
The context
AltLayer's MCP server project was handed over after its backend developers moved to another initiative. The platform needed to let AI agents interact with Web3 dApps and DeFi protocols safely — READ operations free, WRITE operations behind explicit user confirmation — while staying callable by weaker and open-source LLMs that fabricate arguments against large, flat tool surfaces.
Cultivation
What I built
- Took ownership of the handed-over project and built and maintained ~37 protocol MCP servers for Web3/DeFi — Uniswap, SushiSwap, Polymarket, 1inch, Balancer, OKX, DeBank, OpenSea, Spark, EigenLayer, LI.FI, Kalshi, DefiLlama, and others — integrated into AltLLM's multi-model LLM platform.
- Designed the per-protocol server architecture — one focused server per protocol instead of a monolith with a large flat tool surface — keeping each server's tool set small, cohesive, and independently maintainable so tool selection stays tractable for the model.
- Implemented guided tool-calling (`TOOL_GUIDED=true`) — workflow hints, schema-level examples, and `_meta.guided` metadata — so weak and open-source LLMs reliably select and call tools without fabricating arguments.
- Owned the Next.js → stateless Hono migration end-to-end without breaking the public API — MCP protocol handler, auth + OAuth routes, transaction API, sign + OAuth pages as Hono JSX, a framework-agnostic auth library, and the Streamable HTTP transport on @hono/mcp — removing or deprecating stateful implementation details including Prisma, PostgreSQL, and CDP wallet flows.
- Authored the bulk of the protocol coverage end-to-end across lending and staking (Spark, Convex, EigenLayer, Rocket Pool, Marinade, Jito), DEX and aggregators (Balancer V3, 1inch, OKX, ParaSwap, Jupiter), NFT and prediction markets (OpenSea, Kalshi), cross-chain (LI.FI), and data (DefiLlama, CoinGecko, DexPaprika, DeBank).
- Led the READ-first repositioning across 28 protocol tool packages — deferred WRITE / BUILD operations by default so the public surface is safe-by-default, and retired the obsolete API-key-gated build packages.
- Built a ManusAI-like chat UI where tool calls are visible to users, and iterated the interface to support OpenRouter and multiple AI models.
- Integrated wallet transaction flows on Coinbase Developer Platform wallets, later supported the transition toward server wallets, and flipped tools to raw calldata returns so any wallet — CDP, external integrators, or autonomous agents — can consume the catalog.
- Hardened the catalog for multi-chain — RPC fallbacks across every CDP-supported network, deterministic chain string-name resolution, multi-chain token-address enforcement in guided mode, and EVM + Solana parity through CDP embedded wallets.
- Built the LLM tool-validation harness the team uses to gate releases — a Python runner over ~37 protocols and ~176 tools.
Root system
How it works
A stateless Hono API on Vercel serverless fronts ~37 protocol-focused MCP servers over Streamable HTTP. Each protocol gets its own server with a small, cohesive tool set; guided mode injects workflow hints and schema-level examples per tool. READ operations are free; WRITE operations require explicit user confirmation and return raw calldata so any wallet stack can execute. A Python dual-model validation harness exercises the full catalog before release.
Harvest
Outcomes
- ~37 protocol MCP servers and ~176 tools across EVM + Solana, live inside AltLLM's multi-model LLM platform.
- Dual-model LLM validation harness reached ~97% pass rate across ~37 protocols / ~176 MCP tools on altllm-basic.
- Next.js → Hono migration completed with the public API kept stable throughout.
- READ-first repositioning shipped across 28 protocol tool packages — the public surface is safe-by-default.
Reliable agent tooling is an architecture problem, not a prompt problem. Small per-protocol tool surfaces, guided schemas, and READ-first defaults did more for weak-model reliability than any amount of instruction tuning — and an evaluation harness is the only honest release gate.