Give your agents capabilities through tools (function calling). Helps you identify what your agent needs to do, create tool definitions, and attach them to config variations.
This skill guides an agent through giving LaunchDarkly AI configs tool-calling (function-calling) capabilities. Its job is to help identify what an agent needs to do, create tool definitions with JSON Schemas, attach those tools to config variations, and verify the attachment — all through the remotely hosted LaunchDarkly MCP server. The problem it addresses is standardizing how tool schemas are defined once in LaunchDarkly and then reused across providers, so an application can read the stored schema back and convert it to whatever shape each SDK expects.
The workflow is a strict four-step sequence: list existing tools for discovery, create a new tool with `create-ai-tool` (key, description, and raw JSON Schema — not the OpenAI function-calling wrapper), attach it to a variation with `update-ai-config-variation`, and verify with `get-ai-tool` and `get-ai-config`. The skill is emphatic about two operational safeguards: when attaching tools, pass only the `tools` field so that instructions, messages, model, and parameters edited in the LaunchDarkly UI are not silently clobbered by the PATCH; and if a UI-clearing bug is observed, report it rather than working around it by re-sending stale fields. It documents how LaunchDarkly stores the flat `{type, name, description, parameters}` schema once and how application code converts it per provider — OpenAI Chat Completions and Responses API, Anthropic (renaming parameters to input_schema), Bedrock Converse, Gemini, and LangChain/LangGraph/Strands — while stressing that LaunchDarkly never makes the provider call and the tool handlers live in application code.
The target users are developers building LLM agents on LaunchDarkly AI configs who need to define and attach callable tools consistently across frameworks. It is a benign configuration and integration helper that operates through official MCP tooling and read/verify calls, and it does not itself execute any tool behavior.
Identify the capabilities an agent needs, create tool (function-calling) definitions with JSON Schemas in LaunchDarkly, attach them to AI config variations, and verify the attachment via the LaunchDarkly MCP server.
List existing tools first, then create the new tool, attach it to the variation, and finally verify. The skill stresses never stopping after listing alone — always proceed through all four steps.
When attaching tools with update-ai-config-variation, pass only the `tools` field. Bundling instructions, messages, model, or parameters into the same PATCH can silently overwrite edits made in the LaunchDarkly UI.
No. LaunchDarkly stores the tool schema once in a flat shape; your application reads it back, converts it to each provider's expected format, and owns the handler behavior. LaunchDarkly never makes the provider call.
The skill documents converting the stored flat schema for OpenAI Chat Completions and Responses API, Anthropic direct SDK, Bedrock Converse, Gemini (google-genai), and LangChain/LangGraph, among others.
Quick Setup:
.claude/skills/Repository
launchdarkly/agent-skills