@conciergeprotocol/agent

Build the answer function for a Concierge Agent Protocol front door. Give it a model and the company's sources. Pass the function to @conciergeprotocol/server.

The company's tools, keys, and records stay in this process. The visiting agent receives the reply text.

Install

npm install @conciergeprotocol/agent @conciergeprotocol/server

Node.js 22 or newer.

Answer from your own sources

import { createAnswer } from '@conciergeprotocol/agent'
import { createCapServer } from '@conciergeprotocol/server'

const answer = await createAnswer({
  model: 'anthropic/claude-sonnet-4.5',
  instructions: 'You answer for Harbor Goods. Link to the page.',
  sources: [
    { type: 'site', url: 'https://harborgoods.example' },
    {
      type: 'mcp',
      name: 'docs',
      url: 'https://harborgoods.example/mcp',
      headers: { authorization: `Bearer ${process.env.MCP_KEY}` },
    },
  ],
  tools: {
    inventory: inventoryTool,
  },
  maxSteps: 6,
  dailyTokens: 500_000,
})

const cap = await createCapServer({
  agent: { name: 'Harbor Goods', chat: 'https://harborgoods.example/chat' },
  answer,
})
await cap.listen(8088)

A string model is called through the Vercel AI Gateway. Set AI_GATEWAY_API_KEY, or pass a model object from a provider package to call that provider directly.

sources has three kinds:

• site reads a public website at startup: the home page, the pages it links to, then the pages in sitemap.xml, up to maxPages (default 50). The model searches those pages and reads one in full, and each result carries the page URL. Pages on other origins are skipped. Use it when the site has no llms.txt.
• llms fetches an llms.txt file and lets the model read only the pages that file links to.
• mcp connects to an MCP server over HTTP. Its tools are prefixed with name, so a server tool search is docs_search. The header value stays on this side of the door.

tools are AI SDK tools. A tool that changes something is wrapped so a read-only caller cannot run it:

tools: {
  refund: { write: true, tool: refundTool },
}

Write tools are included when executeWrites is true and the CAP grant's scope is write. Otherwise the model does not see them. Holding a change for a person's approval is done by the server: return { confirm } from your own answer function. This package runs a write tool when you set executeWrites.

maxSteps counts model steps, including tool calls. dailyTokens is counted in this process and resets at UTC midnight. A restart clears the count. The server's dailyLimit is a separate cap on messages.

answer.close() closes MCP clients.

What the reply contains

The function returns { reply } with the model's text, or { error: { code: 'failed', message } } when the model fails or the token cap is reached. A model failure is passed to onError, which defaults to console.error, so the owner sees the cause in the server log. Tool names, tool errors, and these instructions are not added to the reply. The prompt tells the model to leave them out.

The visitor message is treated as untrusted. It does not replace the instructions.

License

Apache 2.0