Manage content with AI agents

microfeed supports two interactive agent workflows. Experimental WebMCP lets a browser-based agent read content and help with the draft currently open in a protected dashboard. @microfeed/cli gives a local coding agent the broader API-backed workflow for drafts, media, publishing, and other content tasks.

Use WebMCP for visible drafts

WebMCP is an experimental web platform proposal. A page can expose structured JavaScript tools to an agent in the same browser through document.modelContext. It is not a remote MCP server and does not replace microfeed’s API, OAuth, CLI, OpenAPI document, or llms.txt files.

microfeed activates its WebMCP tools automatically only when both conditions are true:

• the current browser provides the native document.modelContext API; and
• the dashboard is protected by built-in login or Cloudflare Access.

There is no microfeed setting to enable. Public themed pages do not load or run WebMCP code. An unsupported browser performs one feature check and downloads no WebMCP implementation or validation chunk. A supported protected dashboard loads the small tool registry only in that browser session; the server does no polling or background work.

The dashboard exposes tools to list and read Items and Pages, open a new Item or Page editor, and save the visible new or unpublished editor as a draft. A save merges only supplied editorial fields with the editor’s current state and always persists an unpublished result. The tool disappears when the visible content is published, unlisted, deleted, or the default 404 Page. The server rechecks stored and resulting status before accepting every WebMCP write.

WebMCP cannot publish, delete, upload media, manage Site Files or themes, read API keys, change account settings, or configure webhooks. Use the dashboard or the authenticated CLI/API workflow for those operations. Webhook events caused by an accepted draft save use context.origin: "webmcp", so an automation can distinguish them from dashboard and API writes.

Because the browser API remains experimental, tool discovery currently requires a Chrome WebMCP development or origin-trial environment. Browsers without the native API degrade silently.

Use the CLI for broader content tasks

Use @microfeed/cli when a person asks a local coding agent to work across content, upload media, or publish. The CLI wraps the API in task-oriented commands and keeps its browser-granted credentials opaque.

The CLI is an interactive workflow: the person supplies the task and remains available for browser consent and destructive confirmation. A persistent service that reacts asynchronously should instead follow Build webhook endpoints and the build-microfeed-automation skill.

Give the agent an outcome, not a credential

Include the site, the content outcome, and the approval boundary in the prompt:

Use `npx @microfeed/cli` to create a draft item on https://feed.example.com
from article.md. Use `--json`, show me the draft before publishing, and pause
if API access, browser authorization, or destructive confirmation is required.

The command works from any folder. If the agent is already working inside a Git-cloned microfeed source repository whose dependencies are installed, it may use yarn microfeed as a shortcut to the local CLI version. That source repository also provides the canonical manage-microfeed-content skill at .agents/skills/manage-microfeed-content/, which teaches the agent to select commands, preserve credentials, and confirm destructive operations. The published @microfeed/cli package bundles the same skill for agent hosts that distribute skills with npm packages.

Claude Code reads CLAUDE.md in the Git-cloned source repository, which imports the repository guidance and routes content work to that skill. Other compatible coding agents can discover the .agents/skills/ copy directly. You do not need to install or paste the skill into the prompt when the agent is already working inside that source repository.

Keep browser authorization human-controlled

An agent may start npx @microfeed/cli login <site-url>, but the administrator must sign in and approve or deny permissions in the browser. The agent must not click Allow, request a password, or inspect the encrypted credential store or operating-system keychain.

Never paste an API key, OAuth access token, refresh token, or client secret into the prompt. In CI, supply MICROFEED_API_KEY through the secret manager and instruct the agent to reference the variable without reading or printing it.

Choose the operation that matches the request

Ask the agent to inspect current content before writing and to use structured JSON output. Media vocabulary matters:

• an item image is cover art or a thumbnail;
• a media attachment is the item’s one main file and RSS enclosure; and
• standalone media is an uploaded file, such as an inline rich-text image.

The agent should use --image-file, --attachment-file, or media upload accordingly. It should never inspect or print a short-lived upload URL.

For deletion, the agent must report the selected saved-instance name and exact item ID, obtain explicit approval, and then pass that same ID to --confirm. The CLI deliberately has no generic yes flag.

Continue with Manage content with @microfeed/cli for the guided human workflow and the @microfeed/cli reference for the complete command and output contract.