Goldsky Agent

Goldsky Agent

AI-powered tools for the full Goldsky product surface. Build, deploy, and debug Turbo pipelines, Mirror pipelines, Subgraphs, Compose apps, and Edge RPC — from natural-language prompts.

Paste this to get started. The agent installs the CLI, then runs /get-started, which does the rest:

Set up Goldsky. Do this yourself. I only approve login in the browser and tell you what I am trying to do.
Install the Goldsky CLI by running auth-setup/scripts/install.sh with bash. It needs npm and does not download a shell script. If you cannot read that skill yet, install it first with npx skills add goldsky-io/goldsky-agent -y -g -a <agent> (claude-code, cursor, codex, or opencode, or -a '*' if you cannot tell which host you are). Do not curl an installer and do not ask for sudo.
Then run /get-started and follow it.

Installation

Recommended: Universal Skills Installer

npx skills add goldsky-io/goldsky-agent

The installer will prompt you to select your AI agent, or specify directly:

npx skills add goldsky-io/goldsky-agent -a claude-code  # or cursor, opencode, etc.

Works with 30+ AI agents including Claude Code, Cursor, OpenCode, and Codex.

Claude Code (Plugin Marketplace)

/plugin marketplace add goldsky-io/goldsky-agent
/plugin install goldsky@goldsky-agent

Cursor (Local Plugin)

Clone and add to Cursor settings:

git clone https://github.com/goldsky-io/goldsky-agent.git

Then add the path to your Cursor settings (Settings > Cursor Settings > JSON):

{
  "plugins.local": ["/absolute/path/to/goldsky-agent"]
}

Claude Code — load from local directory

git clone https://github.com/goldsky-io/goldsky-agent.git
claude --plugin-dir ./goldsky-agent

Copy skills directly (any tool)

git clone https://github.com/goldsky-io/goldsky-agent.git
cp -r goldsky-agent/skills/* .claude/skills/    # Claude Code
cp -r goldsky-agent/skills/* .cursor/skills/    # Cursor

Quick Start

Just describe what you need in natural language — the right skill is selected automatically.

Use with autonomous or BYO agents

Claude Code and Cursor auto-discover these skills and trigger them from each skill's description. Autonomous or bring-your-own-model agents — Devin, Codex, custom LLM apps — don't implement that convention. Installing isn't enough: npx skills add copies the skills into a skills directory (.claude/skills/ for a single Claude Code target, or the shared .agents/skills/ when targeting several) but writes no entry-point file telling an agent to look there. Bridge that one of two ways:

• Connect the docs MCP (recommended) — any MCP-capable agent can search the docs and discover the skills (exposed as MCP resources) at query time. Devin: add it under mcpServers in .devin/config.json (see Devin's MCP docs). Setup for other clients is under MCP Server below.
• Install the skills, then add an AGENTS.md at your project root so agents that read it (Devin, Codex, Amp, …) know the skills exist — point it at your install path:

# AGENTS.md

This project uses Goldsky. AI-agent skills for building, deploying, and
debugging Goldsky pipelines are installed under `.agents/skills/` (or
`.claude/skills/`). Before a Goldsky task, read the matching
`<skills-dir>/<name>/SKILL.md` and the files it references. For reference
lookups, query the Goldsky docs MCP: https://docs.goldsky.com/mcp

Repository Structure

goldsky-agent/
├── skills/              # All skills (auto-triggered by description matching)
│   ├── turbo-builder/         # Step-by-step pipeline creation wizard
│   ├── turbo-doctor/          # Diagnose and fix Turbo pipeline issues
│   ├── turbo-pipelines/       # YAML config + architecture reference
│   ├── turbo-transforms/      # SQL, TypeScript, dynamic tables
│   ├── turbo-operations/      # Lifecycle commands, monitoring, errors
│   ├── mirror/                # Mirror pipeline deploy, operate, sources/sinks reference
│   ├── mirror-doctor/         # Diagnose and fix Mirror pipelines
│   ├── subgraph-builder/      # Author, build & deploy subgraphs; schema/mappings/manifest
│   ├── subgraph-doctor/       # Diagnose and fix failing/stalled subgraphs
│   ├── subgraph-migrate/      # Guided migration from The Graph
│   ├── compose/               # Compose app scaffolding; references/ + examples/
│   ├── compose-doctor/        # Diagnose and fix Compose apps
│   ├── onchain-automation/    # Cross-product router: detect → decide → execute
│   ├── edge/                  # Managed RPC capabilities, error codes, pricing
│   ├── boost/                 # Free CDN in front of your existing RPC provider
│   ├── feeds/                 # Wallet balances and transfers over the REST API
│   ├── datasets/              # Chain prefixes, dataset types
│   ├── secrets/               # Credential management
│   └── auth-setup/            # CLI installation, login
├── agents/              # `goldsky` subagent definition
├── hooks/               # Optional pre/post deploy automation
│   ├── optional/              # Opt-in Claude Code and Cursor configs
│   └── scripts/               # Validation, secret checking
├── .claude-plugin/      # Claude Code plugin manifest + self-hosted marketplace
└── .cursor-plugin/      # Cursor plugin manifest + logo

How It Works

Skills auto-trigger based on what you describe. Interactive skills guide you through processes, help make decisions, or walk you through multi-step tasks. Reference skills provide syntax lookups, command references, and documentation.

User: "Build me a pipeline for USDC transfers on Base"
  ↓
turbo-builder (skill — auto-triggered)
  ↓ references
turbo-pipelines + datasets + secrets
  ↓
Generated pipeline.yaml + deployment

Skills

Skills are grouped by product. Each group has interactive workflow skills (guided, multi-step) and/or reference skills (lookup-oriented).

Turbo pipelines

Streaming pipelines that index onchain data from 130+ chains into PostgreSQL, ClickHouse, Kafka, S3, and more.

Mirror pipelines

Goldsky's original streaming pipeline product. Prefer Turbo for new pipelines — reach for Mirror only when you need a subgraph entity source, the one thing Turbo can't do.

Subgraphs

Hosted GraphQL APIs over indexed onchain data.

Compose

Offchain-to-onchain TypeScript framework for oracles, keepers, circuit breakers, and cross-chain automation. When you need to run logic and write back onchain — not just read data — Compose is the tool: managed gas, smart wallets, and cron / HTTP / onchain triggers.

Edge (managed RPC)

Globally distributed, low-latency JSON-RPC for EVM chains, built on eRPC — a drop-in replacement for Alchemy / Infura / QuickNode with hedged requests, automatic failover across node vendors, flashblocks, and pay-per-request (x402). Reach for Edge whenever you need a reliable RPC endpoint, not just indexing.

Feeds

Ready-made wallet balances and transfers. Get the key with goldsky feeds key reveal, then call the feeds over HTTP. Do not use the CLI login token.

Cross-cutting

Used across multiple products.

Optional deployment hooks

Deployment hooks are off by default. The CLI already validates duringgoldsky turbo apply, and the skills guide secret setup and post-deployverification. Enable these extra checks only if you want automatic remindersand an additional pre-deploy check. The validation hook runs even when thecommand contains --skip-validation.

When enabled, these hooks run on supported goldsky turbo apply commands:

Hooks require Bash, jq, and the Goldsky CLI. They inspect simple literalgoldsky turbo apply commands, including quoted YAML paths, in the host'sreported working directory. Compound commands, shell expansions, missingprerequisites, and unrecognized secret-list output skip the checks. These hooksare best-effort assistance, not an enforcement boundary. Cursor's post-shellhook has no documented way to add agent context; its reminder may only appearin hook diagnostics. See PRIVACY.md for network and data flows.

Enable for one Claude Code session

From a stable checkout of this repository:

claude --plugin-dir "$PWD" --settings "$(jq --arg root "$PWD" '
  walk(if type == "string"
    then split("${CLAUDE_PLUGIN_ROOT}") | join($root)
    else . end)
' hooks/optional/claude.json)"

This supplies the hook settings for that session without editing your savedsettings. Start Claude without --settings to return to the default behavior.The settings resolve script paths to this checkout, so keep it available.

Enable for a Cursor project

From a stable checkout, print the configuration with absolute script paths:

jq --arg root "$PWD" '
  walk(if type == "string"
    then split("${CURSOR_PLUGIN_ROOT}") | join($root)
    else . end)
' hooks/optional/cursor.json

Add the resulting entries to the target project's .cursor/hooks.json, keeping"version": 1 and preserving any existing hooks. Reload Cursor. Remove theGoldsky entries and reload to disable them. These are project hook settings;they do not depend on plugin-root environment variables at runtime.

See the Claude settings referenceand Cursor hooks reference for host settings.After upgrading from a version with automatic hooks, restart your host andremove any manually registered Goldsky hooks if you want the new default.

Local validation

With Node 22, npm 11.10 or newer, Python 3, Bash, and jq installed:

npm ci
npm run validate:skills
npm test

scripts/validate-skills.mjs checks skill metadata during development and CI;it is not a runtime hook and remains mandatory in CI.

Tests use a stub CLI and temporary files; they do not deploy resources or useGoldsky credentials. Real-host installation and authentication smoke tests arestill required before marketplace submission.

Coverage

The skills cover the full Goldsky product surface:

• Turbo pipelines — 130+ chain sources (EVM, Solana, Bitcoin, Stellar, Sui, NEAR, Starknet); SQL / TypeScript / dynamic table transforms; PostgreSQL, ClickHouse, Kafka, S3, Webhook, S2, SQS, MySQL, Pub/Sub sinks; streaming and job modes; full lifecycle and monitoring
• Mirror pipelines — Subgraph and direct-indexing sources, sinks, lifecycle, plus interactive diagnosis (legacy — prefer Turbo for new pipelines unless you need a subgraph entity source)
• Subgraphs — Author/build/deploy (subgraph-builder: schema design, AssemblyScript mappings, manifest, instant subgraphs, performance, testing, endpoints/tags/webhooks); interactive diagnosis (subgraph-doctor); guided migration from The Graph (subgraph-migrate)
• Compose — compose.yaml manifest, cron / HTTP / onchain triggers, smart wallets, gas sponsorship, TaskContext API, codegen, pricing
• Edge RPC — Capabilities, supported chains, hedged requests, flashblocks, x402, error code lookups
• Boost — Free CDN in front of the RPC provider you already pay for; cacheable reads served from Goldsky, everything else forwarded free
• Feeds — REST API for one wallet's balances and transfers
• Cross-cutting — Authentication, secrets, dataset naming, full CLI reference

MCP Server

The plugin bundles the Goldsky docs MCP server, providing real-time search across Goldsky documentation.

When installed as a plugin, the MCP server starts automatically.

Claude Code

claude mcp add --transport http goldsky-docs https://docs.goldsky.com/mcp

Cursor / VS Code

Add to .cursor/mcp.json or .vscode/mcp.json:

{
  "mcpServers": {
    "goldsky-docs": {
      "type": "http",
      "url": "https://docs.goldsky.com/mcp"
    }
  }
}

Documentation

• Goldsky Docs
• Turbo Pipelines Guide
• CLI Reference

Policies

• License — MIT
• Terms of Use
• Privacy Policy — the plugin collects no data
• Security Policy — report issues to [email protected]