Agents

@hollware/agents runs three agents on a Worker of your own, next to your Hollware:

None of them publishes. Their MCP client refuses every tool that publishes or decides (publish, approve_review, request_changes, unpublish), whatever their token holds. A translation reaches readers only when a person approves its review; an audit is a draft.

Set them up

1. Create a token for them, with no publishing:

npx hollware token create --label agents --scopes content:read,content:write --remote

2. Create a KV namespace for what they remember:

npx wrangler kv namespace create my-cms-agents

3. Make a Worker for them. Its wrangler.jsonc, with the name of your Hollware Worker in service:

{
  "name": "my-cms-agents",
  "main": "src/index.ts",
  "compatibility_date": "2026-10-06",
  "vars": { "HOLLWARE_PUBLIC_URL": "https://cms.example.com" },
  "services": [{ "binding": "HOLLWARE", "service": "my-cms" }],
  "ai": { "binding": "AI" },
  "kv_namespaces": [{ "binding": "AGENTS_KV", "id": "<the id of step 2>" }]
}

And src/index.ts:

import { createHollwareAgents } from '@hollware/agents'
import { WorkerEntrypoint } from 'cloudflare:workers'

const agents = createHollwareAgents({
  auditor: { collection: 'reports', locale: 'en' },
})

export class HollwareAgents extends WorkerEntrypoint<Env> {
  async run() {
    await agents.run(this.env)
  }
}

export default { fetch: agents.fetch, scheduled: agents.scheduled }

4. Give it the token, and deploy:

npx wrangler secret put HOLLWARE_AGENT_TOKEN
npx wrangler deploy

5. Let your Hollware wake them. In the wrangler.jsonc of your Hollware, then deploy it again:

"services": [{ "binding": "AGENTS", "service": "my-cms-agents", "entrypoint": "HollwareAgents" }]

Hollware's cron trigger wakes the agents every ten minutes, so they need no cron trigger of their own: the Workers free plan has five for the account. Each run is one agent, in turn: the reviewer, then the translator, then the auditor, which checks a dozen entries a run and writes its report once a week. One agent at a time keeps a run within the fifty calls a free Worker may make.

If you have cron triggers to spare, give the agents one instead of the binding: scheduled runs them the same way.

The model

The agents use Workers AI, @cf/meta/llama-3.3-70b-instruct-fp8-fast unless you name another. They stop for the day after 100 calls to the model, which keeps them in the free allowance of Workers AI:

createHollwareAgents({ model: '@cf/openai/gpt-oss-120b', dailyLimit: 50, gateway: 'my-gateway' })

gateway sends the calls through an AI Gateway, for its logs and its cache. For another provider, pass a function that takes the messages and returns the answer.

The report

The auditor needs no model. Its report gathers what its runs found in the last week, with the much-read pages that have not changed in six months. It writes the report as a draft of the collection you give it, with the slug audit-<day>, and never publishes it. The collection needs a title and a description in its schema, or a frontmatter function that returns what it does need.

To get the report in Slack as well, set the URL of an incoming webhook as a secret:

npx wrangler secret put AGENTS_WEBHOOK_URL

The auditor reports much-read pages only when the site's analytics are connected: see Search and analytics.