Skip to main content
Field reference: feed_monitors — every field, type, default, and tier for this page.
The Feeds page lets your agent watch the outside world. Point it at RSS feeds or X (Twitter) accounts; when new items appear, the agent can log them, summarise them, react to them, and deliver the result wherever you choose. You can also connect an X account and post tweets directly. Two building blocks:
  1. Monitors — each watches one source (an RSS feed or an X handle/search) on a schedule.
  2. X connection — an OAuth link to one X account, used both to read tweets and to post them.

​
Monitors

Add a monitor with Add RSS feed or Add Twitter monitor. Each monitor has: Use Check now to poll immediately and Recent items to see what the monitor has already seen.

​
Interpreters

Unlike Email and IM Channels — where the pipeline is an owner-only advanced control — feeds run the shared Interpreters stage by default: every monitor has an Interpreter section with an Add Interpreter button. An interpreter recognises what kind of item just arrived and injects a plain-language hint into the agent’s context, so one feed can carry more than one kind of content and the agent still gets the right framing per item. Simple. A feed that’s always the same kind of content doesn’t need interpreters — leave the section empty and describe the feed once in the monitor’s What should the agent do? instructions instead (e.g. “These are blog posts about database internals — summarise each in two sentences.”). Complex. A single feed multiplexes job listings and product announcements. Add two interpreters: one that matches a JSON field for job listings (hint: “extract title, location, salary”), the other regex-matching an announcement marker (hint: “extract product name and launch date”). The agent gets the right framing per item without splitting the source into two separate monitors. See Interpreters for the full matching rules — JSON field presence vs. regex, evaluation order, and how the first match wins.

​
Examples — monitors

Hello world — a reading list. Add an RSS feed for a blog you like. Set On a new item to Off. The monitor logs every new post; you browse them under Recent items. No agent runs, no cost — just a feed reader inside your dashboard. Simple — auto-summaries to your inbox. Add an RSS feed for an industry newsletter. Set On a new item to All, instructions: “Summarise this article in two sentences and note why it matters for an e-commerce business.” Deliver to Inbox. Every new article arrives pre-summarised in your agent’s inbox. Medium — a filtered competitor watch. Add an X monitor for a competitor’s handle. Set On a new item to Classified. Relevance rule: “Only react to posts announcing new products, pricing changes, or partnerships. Ignore memes, replies, and reposts.” Instructions: “Extract the announcement, classify it as product / pricing / partnership, and summarise the implication for us.” Deliver to Email → competitive-intel@yourco.com. Now you get an email only when something strategically relevant happens — the classifier filters out the noise before the agent (and your credits) get involved. Complex — a triaged support-signal pipeline. Add an RSS feed of your product’s public issue tracker plus an X monitor for your brand handle with a keyword filter of down OR broken OR outage. On both, set Classified with a rule that reacts only to genuine problem reports, and instructions: “Determine severity (low/medium/high), draft a holding response, and include the original link.” Deliver to Inbox. Your agent now triages inbound problem signals from two sources into one queue, severity-tagged, with draft responses ready — and you take over from the Inbox to send the real reply. System-pushing — a closed-loop social agent. Connect your X account (below). Add an X monitor for mentions of your brand, Classified to react only to genuine questions (not spam or praise). Instructions: “If this is a clear product question you can answer from the knowledge base, draft a concise public reply (≤280 chars) and post it as a reply to the tweet using the X tools; otherwise summarise it for human review.” With the agent’s X write tools enabled, this becomes a monitored, classified, knowledge-grounded auto-responder — with the classifier and your instructions as the guardrails that keep it from replying to the wrong things.
All mode runs your agent on every new item — costs add up on busy feeds. Prefer Classified with a tight relevance rule for high-volume sources, and start with a longer poll interval until you’ve seen what the feed produces.

​
X (Twitter) connection

Connect one X account via OAuth. Once connected, the agent can read tweets for monitors and — if you give it the write tools — post them. The Post a tweet box lets you compose and send directly from the dashboard, optionally with media.

​
Examples — posting

Simple — manual posting. Connect your account, type into Post a tweet, optionally attach an image, and send. A quick way to post without leaving the dashboard. Complex — agent-authored threads. With the X write tools enabled on the Tools page, your agent can compose multi-tweet threads (a main tweet plus a reply chain), quote-tweet, and attach media — for example, a feed monitor that turns a long article into a 5-tweet summary thread and posts it. Pair this with a Classified monitor so only the right articles trigger a thread.
Posting requires the connected account to have granted write scope, and the agent’s X tools must be enabled. Without write scope the agent can still read for monitors but cannot post.

​
Tips

  • Start with Off. Watch a feed for a day before turning on reactions — you’ll see its real volume and can write a sharper relevance rule.
  • The classifier is cheap; the agent is not. In Classified mode the classifier filters first, so the (more expensive) agent only runs on items worth reacting to.
  • Delivery to Inbox keeps a human in the loop. Use it when the agent should draft but you want to approve before anything goes out.
  • Poll interval is a cost lever. A 5-minute interval on a busy feed reacts fast but polls 288×/day; a 60-minute interval is gentler. Match it to how time-sensitive the source is.