@unbrowse/llamaindex

Unbrowse as native LlamaIndex TS tools. Drop three agent-callable tools into any llamaindex agent so it can resolve and execute live API routes from Unbrowse's marketplace cache.

npm install @unbrowse/llamaindex

Register the tools

unbrowseTools is an array of FunctionTool-shaped objects — register it directly with an agent:

import { unbrowseTools } from '@unbrowse/llamaindex';
import { agent } from '@llamaindex/workflow';
import { openai } from '@llamaindex/openai';

const myAgent = agent({
  llm: openai({ model: 'gpt-4o-mini' }),
  tools: unbrowseTools,
});

const res = await myAgent.run('Find the trending products on example.com');
console.log(res.data.result);

Real FunctionTool instances

If you want first-class FunctionTool objects from your installed llamaindex, inject the class:

import { FunctionTool } from 'llamaindex';
import { createUnbrowseTools } from '@unbrowse/llamaindex';

const tools = createUnbrowseTools({ FunctionTool }); // FunctionTool.from(fn, metadata) under the hood
const myAgent = agent({ tools });

You can also import each tool individually:

import { unbrowseResolveTool, unbrowseExecuteTool, unbrowseSearchTool } from '@unbrowse/llamaindex';

The three tools

| Tool | Input | What it does | |------|-------|--------------| | unbrowse_resolve | { url, intent } | Rank callable API endpoints for an intent against a URL | | unbrowse_execute | { endpoint_id, params? } | Execute a resolved endpoint, return the live body | | unbrowse_search | { query, url? } | One-shot resolve + execute the top endpoint for a query |

Each tool's call returns a JSON string (what an LLM tool loop expects).

Configuration

| Env | Default | Purpose | |-----|---------|---------| | UNBROWSE_API_URL / UNBROWSE_BASE | https://beta-api.unbrowse.ai | Backend base URL | | UNBROWSE_API_KEY | — | Bearer auth (optional) | | UNBROWSE_X_PAYMENT / X_PAYMENT | — | x402 payment header (optional) | | UNBROWSE_DRYRUN | — | Set to 1 for a deterministic offline synthesized result (no network) — used in tests |

Honest scope

This is a thin tool adapter, not a LlamaIndex retriever or vector index. It exposes Unbrowse's resolve and execute operations as agent tools; it does not embed documents, manage a vector store, or implement LlamaIndex's BaseRetriever/QueryEngine interfaces. unbrowse_search is a convenience that resolves then executes the top-ranked endpoint — it depends on a route existing in the marketplace for the given URL/query, and returns an explicit ok: false when none resolves. Quality of results is bounded by what Unbrowse has indexed for the target site.

MIT licensed. See unbrowse.ai/vs/llamaindex.