Skip to content

kaankacar/stellar-ai-guide-mx

Repository files navigation

Hack+ Alebrije | CDMX 2026 — Developer Guide

Mexico has one of the most active SPEI networks in the world, a massive remittance corridor with the US, and millions of people who move money through channels that are slow, expensive, or both. Stellar is built for exactly this, and this hackathon is a chance to build something that actually matters for people here.

This repo is a collection of guides put together by the SDF DevRel team to help you move fast. The goal is simple: every developer at this event should have the same shot at building something real, regardless of whether they have a paid AI subscription, years of Stellar experience, or a high-end laptop.

Start here

About to open Claude Code for the first time? Start with Starter_Prompts.md. It has a ready-to-paste protocol context block, the correct way to describe what you're building so Claude doesn't default to the wrong architecture, and a CLAUDE.md template for Stellar projects.

Building with Mexican peso rails? The regional starter pack (Hackathon_Resources.md) is the fastest path. It has Etherfuse (MXN to CETES via SPEI), AlfredPay (MXN to USDC via SPEI), and BlindPay already wired up as a portable TypeScript library you can drop into any Node project. Before you use Etherfuse, read its section in Dev_Setup_Guide.md. There are several non-obvious gotchas (auth format, customer_id permanence, sandbox simulation) that have cost developers hours.

Don't have a paid AI subscription? Start with Free_AI_Setup.md. The fastest path is NVIDIA Nemotron 3 Super via OpenRouter — a 120B parameter model with a 262K context window, strong on SWE-Bench, completely free. The guide also covers Groq (Llama 3.1 8B at 14,400 req/day), Mistral Codestral (2,000 req/day, code-optimized), Google AI Studio, and a table of providers offering free trial credits ($1-$5). If you'd rather run models locally, it covers six open-source models via Ollama and how to rent a GPU server on RunPod or Vast.ai for a few dollars.

About to write code? Read Dev_Setup_Guide.md first. Five minutes here saves hours later.

Suggested reading order

  1. Starter_Prompts.md before your first Claude Code session
  2. Free_AI_Setup.md if you need a free AI setup
  3. Dev_Setup_Guide.md before writing any code
  4. Hackathon_Resources.md to orient yourself in the Stellar ecosystem
  5. Claude_Code_Guide.md for commands, parallel agents, and browser automation
  6. Recommended_AI_Tools.md to explore what else is available

Starter_Prompts.md

The fastest way to avoid the most common Claude mistake at a hackathon: building a DeFi dApp with Freighter integration when you wanted a self-custodial wallet.

Wallet vs. dApp: A concrete before/after prompt showing how to describe your architecture so Claude builds the right thing on the first attempt. Not a DeFi interface. A self-custodial wallet that generates its own BIP-39 mnemonic, encrypts it with your password, and stores it in localStorage.

Protocol context block: A ready-to-paste block covering SDK version (v14, rpc namespace), Stellar Wallets Kit version, correct USDC issuer for DeFindex compatibility, and network configuration. Paste it at the top of your first message.

Parallel agent prompt template: The exact prompt to spawn three parallel agents (core logic + tests / state + routing / UI components) after your scaffold is done. From the DevRel experiment: same work, 40% less wall-clock time.

CLAUDE.md template: What to put in your project's CLAUDE.md for a Stellar wallet build, including Etherfuse and DeFindex specific notes.

Quick corrective prompts: One-liners for the most common mid-session corrections (wrong wallet type, wrong SDK version, wrong USDC issuer, stuck on DeFindex key, stuck polling sendTransaction).

Dev_Setup_Guide.md

Everything you need before writing code, in one place.

API keys: All protocols that require a key have self-service signup. Etherfuse: https://devnet.etherfuse.com/ramp. DeFindex, AlfredPay, BlindPay and Trustless Work: see Dev_Setup_Guide.md Section 1. Soroswap, Phoenix, Aquarius, and Blend require no key.

Testnet contract addresses: Verified DeFindex, Soroswap, Blend and Trustless Work addresses are included. Aquarius and Reflector Oracle link to their live registries. Phoenix is the only protocol still TBD.

Auth patterns: This is where most developers lose time. Etherfuse uses Authorization: your-api-key with no Bearer prefix. DeFindex uses Authorization: Bearer your-api-key. Trustless Work uses Authorization: x-api-key: your-api-key. They're opposites and none are documented clearly. The guide has correct code snippets for all three.

Testnet asset registry: Testnet USDC has multiple issuers that don't share liquidity; pick the wrong one and swaps silently fail. The guide covers all three (Circle, Blend, and Etherfuse — each separate), plus all five Etherfuse stablebond assets (CETES, USTRY, KTB, CARN, CZERO) with their testnet and mainnet issuer addresses.

Critical gotchas (pulled from 60 build runs):

  • Etherfuse customer_id and bankAccountId are per end-user: generate once per user, store, reuse forever — never per session
  • Etherfuse: a G... address can only be registered to one customer; re-registration fails even in sandbox
  • Etherfuse sandbox orders don't progress automatically: you have to POST to /ramp/order/fiat_received to simulate fiat arriving
  • Etherfuse has a 3-10 second indexing delay after order creation: immediate status queries return 404
  • Etherfuse response envelopes are inconsistent across endpoints: the guide includes a normalizer
  • Soroban: always simulate before sending (simulateTransaction + assembleTransaction before sendTransaction)
  • sendTransaction returns PENDING, not success: you need to poll (or use rpc.Server.pollTransaction)
  • DeFindex: classic Stellar assets must be SAC-deployed before depositing into vaults (all common ones already are)
  • DeFindex: the endpoint is /vault/ not /vaults/, amounts are always arrays, success is HTTP 201
  • Trustless Work: Each role involved must establish a trustline for the asset being used to create the escrow before the escrow is initialized.
  • Trustless Work: The platform role is the only role authorized to modify the escrow at any time. However, once the escrow holds a balance, modifications are restricted exclusively to adding milestones, and no other properties can be updated.
  • Trustless Work: The trustline address must be the issuer address (starting with 'G'), not the contract ID.
  • Trustless Work: The dispute resolver address must be unique.

Section 6 covers known limitations and what to ask your DevRel mentor about, including undocumented API behaviors and known quirks.

Hackathon_Resources.md

Regional starter pack (Mexico + Latin America): https://github.com/ElliotFriend/regional-starter-pack A SvelteKit app with a portable TypeScript anchor library. The /src/lib/anchors/ folder is framework-agnostic; copy it into any TypeScript or Node project and it works without SvelteKit. Covers SEP-1, SEP-6, SEP-10, SEP-12, SEP-24, SEP-31, and SEP-38. Also ships with pre-configured MCP servers for Claude Code (Svelte docs + Etherfuse docs) so Claude understands the stack out of the box.

DeFi reference implementation: https://github.com/kaankacar/stellar-defi-app A full mainnet DeFi dashboard integrating Blend (lending/borrowing with health factor monitoring), Soroswap Aggregator (DEX routing), Phoenix DEX, Aquarius AMM, SDEX, and Reflector Oracle (on-chain USD price feeds). Most useful for seeing how the protocols compose: actual API call shapes, response formats, and how to wire everything together.

AI Integration Series (carstenjacobsen): Four focused Next.js + TypeScript apps built with Claude Code, each paired with a BUILD_REPORT.md:

Community resources:

Videos worth watching:

  • Scoping and Evaluating Your Project (a must-watch before writing any code)
  • Vibe Coding 5 ZK Games in 90 Minutes (live demo of AI-assisted building)
  • The Builder's Guide to AI Prompt Engineering
  • A First Look: Nethermind's SPP (Stellar Private Payments, ZK-based)

All links are in the file.

Claude_Code_Guide.md

Plan mode (Shift+Tab or /plan): Claude reasons through the problem before touching any code. Use it for anything larger than a one-liner. It produces a numbered plan you can edit before it executes.

Parallel agents: The pattern that works: plan mode (20 min) → scaffold agent → 3 parallel agents (core logic + tests / state + routing / UI components) → integration agent → browser smoke test. In the DevRel experiment, this cut wall-clock time by roughly 40% on a 145-minute build. Starter_Prompts.md has the exact prompt to trigger this pattern.

CLAUDE.md: A project-level file that every Claude Code session reads automatically on startup. Put your tech stack, USDC issuer, testnet addresses, and any project-specific gotchas here so you never have to re-explain them.

Browser integration testing: Claude in Chrome can open your running app, click through every user flow, and verify on-chain results without you touching the keyboard. The guide has the full smoke test prompt (create wallet → Friendbot → trustline → lock/unlock → verify balance) and the fix for the oninput event issue in Svelte.

Stellar-specific plugins:

  • stellar-dev:stellar-dev: 8-module playbook covering Soroban contract development (Rust), RPC vs Horizon API reference, frontend + wallet integration (Freighter, Wallets Kit v2, passkeys), classic assets + SAC bridge, 15 documented pitfalls with fixes, security checklist, testing strategy (unit + local Quickstart + testnet), and an ecosystem catalog of DeFi protocols, oracles, and tools. Pre-installed in Claude Code.
  • openzeppelin-skills: three skills for secure Stellar contract development, plus the OZ MCP server for AI-assisted contract generation. Install with /plugin marketplace add OpenZeppelin/openzeppelin-skills.

Full slash command reference, keyboard shortcuts, and CLI flags are in the file.

Recommended_AI_Tools.md

Stellar-native tools:

  • Stella: SDF's official AI assistant for Stellar dev questions. Yellow chat icon at https://developers.stellar.org, also #stella-help on Discord.
  • llms.txt: machine-readable Stellar docs digest for feeding into any LLM. https://developers.stellar.org/llms.txt
  • stellar-dev skill: Claude Code playbook for Soroban, SDKs, RPC, wallet integration, passkeys, and security patterns.
  • OpenZeppelin on Stellar: audited contract library, Contracts Wizard (visual configurator that generates Soroban code), Contracts MCP server, Relayer, Monitor, and Soroban Security Detectors SDK. https://www.openzeppelin.com/networks/stellar
  • x402: per-request HTTP payment protocol for AI agents. Client hits a paywalled endpoint, receives a 402, signs a Soroban auth entry, and retries with payment. Lets AI agents pay for API calls autonomously.
  • Stellar MCP Server (kalepail): exposes Stellar wallet, token, and contract tools to Claude via Cloudflare Workers.
  • XDR MCP (leighmcculloch): decodes and encodes Stellar XDR to/from JSON for AI agents.

Coding assistants with free tiers:

Tool Free tier Best for
Claude Code Free via Ollama (see Free_AI_Setup.md) Agentic terminal coding, full repo context
Continue Fully free VS Code/JetBrains, any local or cloud model
Aider Fully free Terminal + Git, model-agnostic
Cursor 2,000 completions/mo VS Code-like IDE with AI built in
GitHub Copilot 2,000 completions + 50 chats/mo In-editor suggestions
Jules (Google) 15 tasks/day Autonomous GitHub-integrated agent
Windsurf 25 prompt credits/mo Agentic IDE for large codebases

Multi-agent frameworks (for teams building AI-native products): LangGraph, CrewAI, AutoGen (AG2), OpenAI Agents SDK.

Rapid prototyping: Bolt.new (1M tokens/mo, full-stack from a single prompt), v0 by Vercel (React/Next.js UI from text), Google AI Studio (fully free, 1M context window).

Free_AI_Setup.md

Best free cloud option: NVIDIA Nemotron 3 Super via OpenRouter — 120B params, 262K context window, $0/token, strong on SWE-Bench. Works directly with Claude Code (claude --model nvidia/nemotron-3-super-120b-a12b:free). No credit card required.

Other free cloud APIs (no credit card): Groq (Llama 3.1 8B at 14,400 req/day — fastest), Mistral Codestral (2,000 req/day, code-optimized), Google AI Studio (Gemini 2.5 Flash, 1M ctx), Cerebras, GitHub Models. All have Claude Code config snippets in the guide.

Free trial credits ($1-$5): SambaNova ($5 for 3 months), Scaleway (1M free tokens), Nebius, Hyperbolic, Fireworks. Enough for a full hackathon weekend.

Open-source models you can run locally:

Model Best for Min. VRAM
Qwen2.5-Coder-7B Best all-round coding 8 GB
Llama 3.1 8B Best Claude Code compatibility (tool-calling) 10 GB
Phi-4 (14B) Low-end hardware, fast responses 8 GB
DeepSeek-Coder-V2-Lite Algorithms, competitive coding 16 GB
Codestral (22B) Inline autocomplete 16 GB (quantized)
StarCoder2-7B Broad language support, open-source friendly 14 GB

All of these run through Ollama, which installs in one command. The guide walks through the full setup in 8 steps.

No GPU? Rent a GPU server on RunPod ($0.20/hr for an RTX 4090) or Vast.ai ($0.15/hr). SSH-tunnel the port back and Claude Code sees it as if it were local. A full hackathon weekend costs $10-20.

SDF DevRel, March 2026

About

Developer resource guide for Hack+ Alebrije | CDMX 2026 — AI setup, anchor integrations (Etherfuse, AlfredPay, BlindPay), Soroban gotchas, and reference implementations

Resources

Stars

Watchers

Forks

Releases

No releases published

Packages

 
 
 

Contributors