You are probably already using AI to write code. Claude Code, Cursor, Kiro, GitHub Copilot, or some combination. You prompt, you get functions, you move fast. But then you need to build a search application and the workflow fractures. You leave your IDE to read OpenSearch documentation. You switch to the AWS console to provision a domain. You open a terminal to configure indices and pipelines. You write client code to connect everything. Each step requires a different tool, a different mental model, and a different tab.
Your AI agent can generate the code for each piece. What your agent cannot do (until now) is make the architectural decisions, execute the configuration, and verify that the whole system works together. The agent autocompletes functions but has no idea how to actually stand up a search application end to end.
OpenSearch Agent Skills change that. Agent Skills are executable workflows, packaged as SKILL.md files, that give your AI coding agent the domain expertise to build, configure, and deploy OpenSearch applications from natural language. Not documentation. Not code snippets. Actual capabilities your agent can run against your infrastructure.
What Ships Today
The OpenSearch project ships skills covering the workflows where context-switching hurts most: search application development, log and trace investigation, document processing, and cloud deployment to Amazon OpenSearch Service and Serverless.
The Search skill (opensearch-launchpad) builds complete search applications from a single prompt. Tell your agent "build a hybrid search app for my product catalog" and the skill handles index creation, ingest pipeline configuration, ML model registration with Amazon Bedrock, retrieval strategy selection (BM25, semantic, hybrid, or agentic), and a working search UI. The skill supports four retrieval strategies and knows when to recommend each one. Ask for a search bar and it configures a flow agent. Ask for a chatbot and it sets up a conversational agent with memory. You stay in your IDE the entire time.
The Logs skill (log-analytics and trace-analytics) brings the same pattern to observability. Instead of manually writing PPL queries to hunt for error patterns or navigating trace waterfalls to find the slow span, you describe the symptom: "why is my service returning 500s?" The skill queries your log data, identifies error patterns and volume anomalies, investigates distributed traces to find failing spans and service dependencies, and correlates logs to traces by traceId. Root cause analysis that used to require expertise in PPL syntax and trace data structures now requires a sentence.
The Cloud skills (aws-setup, aiven-setup) handle deployment. Once your search application works locally in Docker, the cloud skill deploys the same configuration to Amazon OpenSearch Service or Serverless. No manual console work, no translating local settings into production infrastructure. The Ingest skills cover two paths: document-processing handles local document ingestion via Docling (parse PDFs, HTML, office documents into indexable chunks), while managed-ingestion-service configures Amazon OpenSearch Ingestion pipelines with Automatic Semantic Enrichment for production scale. Separately, the OpenSearch Migration Assistant handles Solr-to-OpenSearch migrations as its own packaged tool, though Agent Skills may absorb that workflow in a future release.
What "Build a Search App" Actually Looks Like
Here is the search skill in action. You open Claude Code (or Cursor, or Kiro) and type: "Build a hybrid search application for my product documentation." The agent activates the opensearch-launchpad skill and moves through five phases.
First, data strategy. The agent asks what data you want to search and offers to work with your files, pull from a URL, or use a built-in sample dataset for prototyping. Second, retrieval strategy. Since you said hybrid, the agent knows to configure both BM25 keyword matching and semantic vector search with a neural pipeline. Third, architecture planning. The agent presents the full plan: index name, field mappings, embedding model from Amazon Bedrock, ingest pipeline configuration, search pipeline with hybrid scoring, and a React frontend. You review and approve.
Fourth, execution. The agent spins up a local OpenSearch cluster in Docker, creates the index, registers the Bedrock model, builds the ingest and search pipelines, indexes your data, and launches the UI. You have a working hybrid search application running locally. Fifth, deployment. The agent offers to deploy the exact same configuration to Amazon OpenSearch Service. Not a re-implementation. The same domain, same index, same pipelines, same model connector. Local-to-cloud with no translation layer.
The whole sequence takes minutes. The agent made the architectural decisions (which retrieval strategy, which embedding model, how to wire the pipeline), executed the configuration, verified the results, and deployed. You never left your IDE. You never opened the AWS console. You never read a documentation page about ingest processor syntax.
How Skills Work (The man Pages Analogy)
Think of Agent Skills as pre-read man pages with executable authority. Your AI agent already knows how to generate code. What the agent lacks is the domain-specific judgment about what to configure, in what order, with what settings. Agent Skills supply that judgment as structured SKILL.md files: step-by-step workflows, reference documentation, and executable scripts bundled together.
Skills load on demand. You can install the full collection without bloating your agent's context window. When you say "build a hybrid search app," the agent activates only the matching skill, follows its workflow, and calls the right OpenSearch APIs. When you say "investigate my 500 errors," a different skill activates. Each skill is small enough to fit in a tight context window but complete enough to handle real end-to-end workflows.
The format is agent-agnostic. Agent Skills were developed by Anthropic as a lightweight, open specification. Any agent that supports the Agent Skills protocol can use them: Claude Code, Cursor, Kiro, VS Code, GitHub Copilot, Codex. You install once and every compatible agent in your workflow gains the same capabilities.
Getting Started
Installation is one command. Run npx skills add opensearch-project/opensearch-agent-skills and your agent gains access to the full skill library. Or install specific skills: npx skills add opensearch-project/opensearch-agent-skills@opensearch-launchpad --full-depth for search, @log-analytics for log investigation, @trace-analytics for distributed traces, @aws-setup for cloud deployment. Prerequisites are Python 3.11+, uv, and Docker running locally. AWS credentials are optional (needed only for deploying to OpenSearch Service).
Once installed, express your intent. "I want to build a semantic search app with OpenSearch." The agent reads the skill instructions, runs the scripts, and handles the rest. No MCP server required. No additional tooling. The skills are files your agent reads and executes.
This Is Open Source. Build With Us.
The skills repository lives at github.com/opensearch-project/opensearch-agent-skills. It is open source, Apache 2.0 licensed, and accepting contributions. The launch skills cover search, observability, ingest, and cloud deployment. The roadmap includes dashboards, security configuration, migration workflows, and more. But the most interesting skills will come from people solving real problems in production.
If you have built a workflow that reliably configures OpenSearch for a specific use case, that workflow is a candidate for a skill. Package the steps, the reference docs, and the scripts into a SKILL.md file and submit a PR. Every skill you contribute means every developer using an AI coding agent gains that expertise automatically. You solve a problem once. The entire community benefits.
