skill
superlight-exa-skill
AI-powered semantic web search skill for AI coding assistants via Exa API. Minimal tokens, maximum discovery. Multi-key rotation with 429 failover.
Set up
Skills are installed by placing their folder where your agent loads skills.
The download comes from github.com/edxeth/superlight-exa-skill itself, not from our server, so what you get is exactly what its author published.
Install instructions come from the README
README
synced Oct 23, 2026Superlight Exa Skill
AI-powered semantic web search via the Exa API. A superlight agent skill for AI coding assistants — minimal tokens, maximum discovery. Supports multiple API keys with round-robin rotation and automatic 429 failover.
Features
- Semantic web search — AI-powered embeddings-based search for relevant content
- Fact verification — Clear doubts and verify information with authoritative sources
- Debugging support — Search for error solutions, stack traces, and troubleshooting guides
- AI answers — Get LLM-generated answers with citations
- Code search — Developer-focused search across GitHub, StackOverflow, etc.
- Similar pages — Find pages similar to a reference URL
- Token-efficient — Minimal context overhead with progressive disclosure
- Multi-key rotation — Round-robin distribution with automatic 429 failover
Why Use This Over Exa MCP?
| Aspect | MCP Server | This Skill |
|---|---|---|
| Context cost | ~700+ tokens always¹ | ~82 tokens always + ~630 on-demand |
| Tool schemas | Always in context | None (progressive disclosure) |
| Setup | Requires MCP configuration | Drop-in skill directory |
| Dependencies | Node.js runtime | bash, curl, jq (Linux/macOS) |
¹ Estimated from multi-tool MCP measurements (~14k tokens for 20 tools). Source
Best for: Users who need web search on-demand without persistent context overhead.
Token Budget
Uses Claude's progressive disclosure architecture:
| Level | When Loaded | Content | Tokens |
|---|---|---|---|
| Metadata | Always (startup) | Skill description | ~82 |
| Instructions | When triggered | SKILL.md protocol | ~630 |
| Resources | As needed | troubleshooting.md | ~632 |
Token counts measured with claudetokenizer.com (Claude Sonnet 4.5)
Installation
Quick Install (Recommended)
npx skills add edxeth/superlight-exa-skill
The installer will prompt you to select which agents to install to (Claude Code, Cursor, OpenCode, Codex, Antigravity, etc.).
Manual Installation
Clone directly to your agent's skills directory:
# Claude Code
git clone https://github.com/edxeth/superlight-exa-skill.git ~/.claude/skills/exa
# OpenCode
git clone https://github.com/edxeth/superlight-exa-skill.git ~/.opencode/skill/exa
Directory structure:
exa/
├── SKILL.md
├── reference/
│ └── troubleshooting.md
└── scripts/
└── exa.sh
Usage
The skill triggers automatically when searching for web content:
"Find the latest research on LLM context windows"
"What is SpaceX's current valuation?"
"Find similar pages to this GitHub repo"
"Search for React useCallback TypeScript examples"
Manual Invocation
# Web search
./scripts/exa.sh search "latest AI research" 5 "research paper"
# Get page contents
./scripts/exa.sh contents "https://arxiv.org/abs/2307.06435"
# Find similar pages
./scripts/exa.sh similar "https://github.com/anthropics/anthropic-cookbook" 5
# AI answer with citations
./scripts/exa.sh answer "What is the current valuation of SpaceX?"
# Code-focused search
./scripts/exa.sh code "React useCallback hook TypeScript examples"
API Endpoints
Uses Exa REST API:
| Endpoint | Purpose | Rate Limit |
|---|---|---|
POST /search | Semantic web search | 5 QPS |
POST /contents | Get page contents | 50 QPS |
POST /findSimilar | Find similar pages | 5 QPS |
POST /answer | AI answer with citations | 5 QPS |
Configuration
API key is required.
# Single API key
export EXA_API_KEY="your-key-here"
# Multiple API keys for load distribution
export EXA_API_KEY="key1,key2,key3"
When multiple keys are provided (comma-separated), the script rotates through them in round-robin order, ensuring even distribution of requests. If a key hits rate limits (429), the script automatically fails over to the next key and retries, only failing after all keys are exhausted across multiple retry rounds.
Get an API key at dashboard.exa.ai.
Requirements
- Platforms: Linux, macOS
- Dependencies: bash, curl, jq
Skill Metadata
name: exa
description: Searches the web using Exa's neural embeddings-based search API. Use when needing real-time web information to answer questions, verify facts, debug issues, find code examples, research topics, or clear doubts with authoritative sources.
License
MIT License