@sharelyai/tools

First-party Sharely tool definitions — the LLM-facing contracts for the 7 platform tools (search_knowledge, semantic_search, get_knowledge_item, list_taxonomies, get_taxonomy_knowledge, get_workspace_stats, list_roles).

The same definition objects are consumed by sharelyai-be's hosted agent runtime, so the JSON schema your LLM sees is identical whether the agent runs Sharely-hosted, Vercel AI–driven, or as a raw Handler.

Install

npm i @sharelyai/tools @sharelyai/protocol

What's shipped

• definitions — typed ToolDefinition[] for all 7 tools.
• Per-tool exports: searchKnowledgeDefinition, semanticSearchDefinition, getKnowledgeItemDefinition, listTaxonomiesDefinition, getTaxonomyKnowledgeDefinition, getWorkspaceStatsDefinition, listRolesDefinition.
• getToolDefinitions() / getDefinitionByName(name) — lookup helpers.
• createTools(executors) — produces a Tool[] for an executor registry you supply. Tools without a registered executor return { error: "Tool ... has no executor wired" }.
• createPlatformExecutors(api) — platform-backed executors for the tools that have a Backplane endpoint today. Currently semantic_search, backed by @sharelyai/api's rag().
• executeTool(name, input, ctx, executors) — dispatches by name.

The execute layer

The upstream sharelyai-be execute functions hit Prisma / Pinecone / OpenAI embeddings directly — those can't ship in a public SDK. Instead:

• semantic_search works out of the box — createPlatformExecutors(api) backs it with @sharelyai/api's rag() (embedding + vector retrieval).
• The other 6 tools (search_knowledge, get_knowledge_item, list_taxonomies, get_taxonomy_knowledge, get_workspace_stats, list_roles) have no Backplane endpoint yet — plug your own executor. search_knowledge in particular is keyword/ILIKE search and is not backed by rag() (that would silently turn a keyword tool into semantic search).

import { createTools, createPlatformExecutors } from '@sharelyai/tools';

const tools = createTools({
  ...createPlatformExecutors(api), // semantic_search, ready to use
  search_knowledge: async (input, ctx) => {
    // bring your own
    return { output: { totalResults: 0, results: [] } };
  },
});