Lloyaldocs
lloyal.ai GitHub
Abilities

First-party abilities

Four sources of evidence, signed and ready: Wikipedia, the open web, a folder of your documents, and whatever a reader attaches.

On this page
Ability Install Lets agents Needs Settings
wikipedia lloyal/wikipedia Search and read Wikipedia, no key — none
web lloyal/web Search the open web and read pages, admitting only the passages that answer reranker tavilyKey (optional, secret)
corpus lloyal/corpus Search, grep and read a folder of your own documents reranker corpusPath (required)
documents lloyal/documents Search, read and look at the files a reader attaches reranker, vision none

The basic template ships wikipedia; the research template ships web, corpus and documents.

Install one

Terminal
npx lloyal-ai install lloyal/web

The install verifies the signed package before it writes anything, vendors it into vendor/, and — if it needs a service your harness.yml does not name — offers to add the line. Then enable it in src/app.ts:

src/app.ts
import { createWebAbility } from "@lloyal-labs/web-ability";

export const abilities = [createWikipediaAbility, createWebAbility];

Publish and install has what the install checks.

wikipedia

Answering questions about established facts, historical events, biographies, scientific concepts, geography, or other general-knowledge topics covered by Wikipedia.

Tools: wikipedia_search, wikipedia_fetch. Wikipedia's public API, no key. It has no reranker: results are a capped list and the model reads what comes back. It carries its own guard, title_dedup, so one agent does not fetch the same article twice.

web

Gathering evidence from the open web — verifying current claims, retrieving primary sources from URLs, surveying official documentation and authoritative discussion.

Tools: web_search, fetch_page. fetch_page chunks a page on its headings and admits the best passages verbatim within a token budget — see Retrieval. With a tavilyKey it searches through Tavily; without one it uses a paced keyless search.

harness.yml
abilities:
  web:
    tavilyKey: ""        # better set from the settings pane or the environment than committed

corpus

Investigating a local document corpus — finding occurrences of terms, reading specific files at line offsets, semantic retrieval over indexed corpus content.

Tools: grep, read_file, search. search narrows a folder of Markdown lexically, then the reranker judges the candidates; read_file returns only the lines an agent has not already read. When corpusPath changes under a live run, the new index is built beside the old one, which keeps serving until it is ready.

harness.yml
abilities:
  corpus:
    corpusPath: ./docs

It is not enabled until corpusPath is set.

documents

Reading the documents attached to this conversation — PDFs, papers, reports: finding passages, quoting sections with page numbers, and checking a table or figure on a specific page.

Tools: search_documents, read_document, view_page. See Attachments and documents for how each spends the model's context, and how citations reach a page.

↑
Search