Concierge Agent Protocol for JavaScript

Packages for both sides of a CAP front door. @conciergeprotocol/client calls a door. @conciergeprotocol/mcp exposes that client as three tools, so Claude, Cursor, and other MCP hosts can ask a company. @conciergeprotocol/server runs a door from one answer function. @conciergeprotocol/agent builds that function from a model, an llms.txt file, and MCP tools.

The protocol is at conciergeprotocol.org. A company publishes agent.txt. It does not publish an MCP server for visitors. An MCP server, when there is one, stays behind the door.

Install

npm install @conciergeprotocol/client

The MCP server is a command. Install it, or run it with npx as in the block below.

npm install @conciergeprotocol/mcp

Client

import { createClient } from '@conciergeprotocol/client'

const cap = createClient({ from: 'agent:my-assistant' })

const answer = await cap.ask('example.com', 'Do you have a free plan?')
console.log(answer.reply)

A follow-up sends only the new question:

const thread = cap.thread('example.com')
await thread.ask('What do you sell?')
const price = await thread.ask('How much is the cheapest paid plan?')

discover returns null when a site has no front door. ask throws NoDoorError for the same case, which is the signal to browse the site instead.

ask waits up to two minutes. If the company is still answering, the message comes back with status: 'working' and no reply. get reads it again.

Account questions use login. The client opens the company's page, and finishLogin stores the token for the next ask. The token stays in the token store you pass to createClient.

MCP

Cursor, Claude Desktop, and VS Code use the same block. In VS Code it goes in .vscode/mcp.json. In Cursor and Claude Desktop it goes in the MCP settings file.

{
  "mcpServers": {
    "cap": {
      "command": "npx",
      "args": ["-y", "@conciergeprotocol/mcp"],
      "env": { "CAP_FROM": "agent:my-assistant" }
    }
  }
}

CAP_FROM defaults to agent:cap-mcp. The tools are cap_discover, cap_ask, and cap_check. The server is listed in the MCP registry as io.github.conciergeprotocol/mcp. cap_ask waits 60 seconds. If the company is still answering, it returns the message id and the model calls cap_check instead of guessing.

The same install block is in packages/mcp/README.md and listing/install.md. Directory copy and the logo are in listing.

Run a door

answer is the company's own function. The server keeps the thread, hides one caller's reply from another, and enforces the daily message cap.

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: 'llms', url: 'https://harborgoods.example/llms.txt' }],
})

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

await cap.listen(8088)

A company that already has a bot can skip the agent package and pass its own answer. npx cap-check http://127.0.0.1:8088 reads agent.txt, asks one question, and checks that a second caller cannot read the reply.

Release

Add a changeset, then merge to main:

bun run changeset

Package Checks runs on every push. When it passes on main, the publish workflow opens a version pull request, or publishes both packages and tags the release. The two packages stay on one version. When the npm publish succeeds, the same workflow lists the server in the MCP registry. The repository needs an NPM_TOKEN secret.

License

Apache 2.0

Concierge Agent Protocol for JavaScript