X (x.com) for AI Agents

The x.com config covers 23 actions across 9 screens, and it is built for reading and research. Search is the anchor: search lands on chronological Latest results, set_search_tab re-runs the same query under Top, People, Media, Photos, or Videos, and apply_advanced_query composes X's full operator grammar (from:, since:, until:, min_faves:, filter:media, lang:, OR, exact phrases) into a single query string. list_available_filters reports what applies in the current context, because filters are dynamic, not fixed.

From any result, the agent can go deeper: open_post reads a full thread with replies, load_more_replies expands it, open_quotes and open_reposts map who amplified a post, and open_profile plus set_profile_tab walk an account's posts, replies, highlights, followers, and following. get_trending_topics and set_explore_tab cover what's trending by category (for_you, trending, news, sports, entertainment), and open_list and open_community follow curated List and Community feeds by URL.

These come straight from the mapped action behavior, and they matter for correct results.

**Permalinks end in /analytics.** The permalink field on each post is a relative path like /handle/status/id/analytics. To open one with open_post, prepend https://x.com and strip the trailing /analytics. Only use ids that came back in results; never fabricate a status id.

**Top and Latest are not supersets of each other.** Latest is chronological and dedups likely spam; Top applies opaque ranking. For real topic coverage, sweep both tabs and union by permalink. Also, spaces in queries must be %20-encoded, never +.

**Infinite scroll has a ceiling.** load_more scrolls a virtualized feed and appends rows, but it only reaches so far back. For deep history, deep_paginate appends a max_id: snowflake cursor to the query (smallest post id seen so far, minus 1) and walks backwards in time, call after call.

**Some surfaces just don't exist or don't read.** X has no public likers page for a post (there is open_reposts, but no likers equivalent), a profile's likes tab is usually private and redirects back to the profile, and the media tab is an image grid with no post text.

X puts nearly everything behind a login, and the mapping reflects that: every core action requires an authenticated session. The config records the login page (https://x.com/login) and the two cookies X keys a session on, auth_token and ct0. Once your session is captured through Rindler's login capture flow, it persists across the agent's calls, so the agent searches, reads threads, and browses profiles as your signed-in account without re-authenticating each time. A test-only login action (username, password, otp) also exists for the hosted handoff flow used in development.

Every read lands as records in data.listings, and the mapped screens tell you what shape to expect. Search and timeline screens (search_results, list_timeline, community_timeline, post_thread) return posts: permalink, profile_url, timestamp, full_text, and engagement_summary with raw counts. People screens (search_people, profile_people, post_reposts) return accounts: name, profile_url, and a summary combining display name, handle, and bio. The explore_trends screen returns trend rows with a name and an optional search_url.

Three settle actions do the reading: list_results for the current search, timeline, or thread, list_user_posts for a profile timeline, and list_people for account results. Each read is capped at a handful of records, so the working loop is read, dedup by permalink, load_more, and stop when the deduped set stops growing.