SDK Tutorials

SDK Tutorials — Real Problems, Real Solutions

These tutorials don’t waste your time with “hello world” demos. Each one solves a problem that costs real money, real engineer-hours, or both. They’re short, opinionated, and brutally focused on outcomes.

Need something demoable today?

Start with the runnable SDK example tracks below before you go deep into the full tutorials.

Business need	Best example track
Customer support triage	business_case_showcase + session_aware_support
Executive update drafting	business_case_showcase
Budget-sensitive engineering help	cost_aware_review
Multi-step analyst work	parallel_research + multi_agent_pipeline

Verified SDK proof matrix

These examples are not aspirational — they are backed by fresh local end-to-end verification.

Target	Coverage proof
Rust	34/34 passed
Node.js	30/30 passed
Python	28/28 passed
WASM	17/17 passed

The Five Tutorials

#	Tutorial	Problem Solved	Measured Win
1	Cost-Aware Code Review	LLM bills from over-using flagship models on trivial work	60–90% cost cut vs. always-Opus baseline
2	Parallel Research Pipeline	Sequential LLM calls that take 30s when they could take 6s	5× throughput on independent queries
3	Multi-Agent Documentation	One model trying to do 5 jobs poorly	Specialist fork pattern with shared context
4	Session-Aware Support Bot	Bots that forget who the user is and re-ask everything	FTS5 search across every prior conversation
5	Safe SQL Agent (Custom Tool)	Agents that can read anything — including your prod DB	Allow-list + approval gate for destructive ops

Who These Are For

If any of these apply, read on:

┌────────────────────────────────────────────────────────────────┐
│  You are building an LLM product and…                         │
│                                                                 │
│   ✓  Your monthly LLM bill just doubled                        │
│   ✓  Users complain the bot is "slow"                          │
│   ✓  You need to mix models (cheap + expensive) in one flow    │
│   ✓  You need tool-use with human-in-the-loop approval         │
│   ✓  You need to embed an agent in an existing app (not CLI)   │
└────────────────────────────────────────────────────────────────┘

Architecture at a Glance

The three EdgeCrab SDKs are thin language wrappers around one Rust engine — the same engine that powers the CLI. No REST, no IPC, no Python-subprocess-spawns-Rust hacks.

                ┌─────────────────────────────────┐
                │    edgecrab-sdk-core (Rust)     │
                │  SdkAgent │ ToolRegistry │ DB   │
                │     ReAct loop │ streaming      │
                └──┬────────────┬────────────┬────┘
                   │            │            │
         ┌─────────┘            │            └─────────┐
         │                      │                      │
    ┌────┴─────┐         ┌──────┴─────┐         ┌──────┴─────┐
    │   Rust   │         │   Python   │         │  Node.js   │
    │ edgecrab │         │  edgecrab  │         │  edgecrab  │
    │   -sdk   │         │  (PyO3)    │         │  (napi-rs) │
    └──────────┘         └────────────┘         └────────────┘
         │                      │                      │
    in-process             in-process             in-process
    zero-copy              zero-copy              zero-copy

This matters because:

• No network hop → p50 latency matches CLI.
• Shared session DB → a Python script and a Node.js dashboard can read the same conversation history.
• Same toolset → any tool that works in the CLI works in every SDK.

Conventions Used

• All code samples are tested and known-working on EdgeCrab SDK v0.6+.
• Every tutorial has a verification block you can paste into a terminal.
• Every tutorial has a cost/latency table showing the win.
• We recommend starting with Cost-Aware Code Review — it sets up patterns used throughout.

Prerequisites

# 1. Install EdgeCrab (any SDK)
# Rust:
cargo add edgecrab-sdk tokio --features tokio/full

# Python:
pip install edgecrab

# Node.js:
npm install edgecrab

# 2. Set one of these API keys in your shell
export OPENAI_API_KEY=sk-...           # OpenAI
export GITHUB_TOKEN=ghp_...            # GitHub Copilot (gpt-5-mini and related catalog access)
export OPENAI_BASE_URL=http://localhost:11434/v1  # Local Ollama / compatible endpoint

Running the Tutorials

Every tutorial works with any model in the catalog. We recommend copilot/gpt-5-mini for the cost-sensitive examples and openai/gpt-4o or openai/gpt-5 for the reasoning-heavy ones.

To swap models in any example:

// Rust
let agent = SdkAgent::new("copilot/gpt-5-mini")?;

# Python
agent = Agent("copilot/gpt-5-mini")

Node.js

const agent = new Agent({ model: 'copilot/gpt-5-mini' });

---

Ready? Start with Tutorial 1 — Cost-Aware Code Review →