A library of expert prompts is only useful if something picks the right one, and that is what `godle_match_capability` does — it takes the task in plain language and returns ranked candidates with relevance scores, so the agent does not have to know the catalogue. The split between layers is honest about where the cost is: reading roles and templates is free and unauthenticated, executing them statefully is not.
A prompt-orchestration library exposed as agent tools. It holds structured expert prompt templates across 24 professional categories — engineering, data, product, marketing, finance, legal, HR and more — each with input and output schemas, eval rubrics and model routing hints. The tools let an agent find the right role for a task rather than being told which one to use.
- `godle_match_capability` ranks roles and templates against a natural-language intent, with optional constraints for domain, riskLevel, modelHint and maxCostUsd
- `godle_execute_task` runs a template by `roleId` and `templateKey` with your input, and can stream over SSE
- `godle_list_roles` lists the roles, filterable by category
- `godle_compose_workflow` builds a multi-step workflow from the library
- The underlying data is public JSON: 185 roles, 1,741 eval rubrics, 24 categories, 12 multi-step workflows and 8 agent team packs
The read layer needs no key at all — the role, template, workflow and eval endpoints are static JSON with CORS enabled, served under `https://godle.app/api/v3/`. The tools that hold state — capability matching, task execution with sessions and streaming, eval-driven retries — sit on a second layer that requires an API key. The MCP manifest lives at `https://godle.app/.well-known/mcp.json`. The SDK itself (`godle-sdk`, version 3.0.0) is zero-dependency and runs in a browser or Node.