Web search

Give your app's AI agent web search and page reading, without paying a search API. mybase asks its own SearXNG (several search engines at once) and turns pages into markdown.

Turn web search on from the project's page or with mybase search enable <project>. Your server authenticates with a mybase API key that has search:read for the project: mybase token create shop-agent --project shop --perm search:read. Keep it server-side, as MYBASE_API_KEY.

Search and readCopy Anchor Link

import { createSearchClient } from "@mybase-dev/sdk/search";

const web = createSearchClient({
  apiUrl: "https://mybase.example.com",
  project: "shop",
  apiKey: process.env.MYBASE_API_KEY!,
});

const { results } = await web.search("bun sql postgres", { limit: 5 });
// [{ title, url, snippet, engines: ["bing", "mojeek"], publishedAt: null }, …]

const page = await web.fetch(results[0]!.url, { maxChars: 10_000 });
// { url, title, markdown, truncated, contentType, cached }

search(query, options) takes limit (1–20, default 10), category (general, news, it for GitHub and Stack Overflow), timeRange (day, week, month, year), language (en, en-US), and page (1–10).

fetch(url, options) reads a public web page and returns its main content as markdown, cut at maxChars (default 20000). Text and JSON come back as they are. Pages on private or local addresses are refused.

Both throw MybaseApiError (with status) when mybase answers with an error: 400 for bad input or a blocked URL, 404 when the service is off, 415 for a page that isn't HTML, text, or JSON, 429 past the per-project limits, 502 when the page or every search engine failed, 503 when SearXNG is down.

Tools for an agentCopy Anchor Link

searchTools(client) returns web_search and web_fetch with a JSON Schema and a run function, for any LLM tool-use loop. run returns text ready for the tool result. Problems the model can work around (a blocked page, no results, rate limits) come back as text starting with Error: instead of throwing.

import Anthropic from "@anthropic-ai/sdk";
import { createSearchClient, searchTools } from "@mybase-dev/sdk/search";

const tools = searchTools(createSearchClient({ apiUrl, project: "shop", apiKey: process.env.MYBASE_API_KEY! }), {
  maxChars: 15_000,
});
const anthropic = new Anthropic();
const messages: Anthropic.MessageParam[] = [{ role: "user", content: "What's new in the latest Bun release?" }];

while (true) {
  const response = await anthropic.messages.create({
    model: "claude-sonnet-5-5",
    max_tokens: 4096,
    tools: tools.map((t) => ({ name: t.name, description: t.description, input_schema: t.inputSchema })),
    messages,
  });
  messages.push({ role: "assistant", content: response.content });
  if (response.stop_reason !== "tool_use") break;
  const results = await Promise.all(
    response.content
      .filter((block) => block.type === "tool_use")
      .map(async (block) => ({
        type: "tool_result" as const,
        tool_use_id: block.id,
        content: await tools.find((t) => t.name === block.name)!.run(block.input),
      })),
  );
  messages.push({ role: "user", content: results });
}

For OpenAI-style APIs, map each tool to { type: "function", function: { name, description, parameters: inputSchema } }.

Good to knowCopy Anchor Link

• Searches are cached for 1 hour and pages for 1 day, so repeating a call is cheap (cached: true).
• Limits: 60 searches and 120 page reads a minute per project.
• Searches go out from the mybase server's IP. If engines rate-limit it, results get thinner for a while; if all of them fail, search throws a 502 saying which ones and why.
• No JavaScript is run on fetched pages. Pages that render in the browser come back mostly empty.