OpenClaw plugin for creating high-quality X posts from current Birdclaw context and user-defined topics.

xsquared is local-first:

• Birdclaw is used for X/Twitter local context and final posting.
• The CLI is written in TypeScript under src/ and compiled to dist/.
• Drafts, trend snapshots, writing-profile snapshots, and rewrite requests live under .xsquared/ in the plugin repo unless XSQUARED_HOME is set.
• Posting to X happens only when the user clicks the dashboard post button or runs the explicit post command.

Quick start:

npm install
npm run build
npm run check
npm run dashboard

Then open the dashboard URL printed by the command. Each source has its own routable tab:

http://127.0.0.1:3888/sources           (most-recent source or empty state)
http://127.0.0.1:3888/sources/<id>      (a specific topic or viral source)
http://127.0.0.1:3888/profile

OpenClaw workflow:

• Create a new topic source for AI agents and run research.
• Generate 5 posts from the AI agents source.
• Rewrite the newest xsquared draft.
• Open the xsquared dashboard.

Sources, the new content model:

• A topic source has an angle + seed notes. Run research to call claude -p with WebSearch/WebFetch and produce a research artifact (summaries + facts + links + cost). Then generate produces 5 LLM-drafted posts grounded in the research.
• A viral source has a filter (e.g. "AI agents"), a Birdclaw resource (home / following / for-you), and a limit. Run viral-fetch to pull the current feed; select 1 or more posts; then generate produces one LLM-drafted reply-style post per selection.
• Both flows fall back to local template generation when the claude CLI is missing or XSQUARED_DISABLE_LLM=1 is set.

CLI:

npm run build
node dist/xsquared.js doctor                                       # checks birdclaw + claude
node dist/xsquared.js sources                                      # list sources
node dist/xsquared.js source-new --kind topic --name "AI agents" --angle "..." --notes "..."
node dist/xsquared.js source-new --kind viral --name "AI feed" --filter "AI agents"
node dist/xsquared.js research <source-id>                         # LLM research, ~30-90s
node dist/xsquared.js viral-fetch <source-id>                      # pull viral feed
node dist/xsquared.js generate <source-id> --count 5               # 5 LLM drafts
node dist/xsquared.js generate <source-id> --selected id1,id2      # viral: only these
node dist/xsquared.js auto-draft [source-id] --max-drafts 3        # fetch, rank, dedupe, draft
node dist/xsquared.js validate-draft <source-id> --text "..."     # grounding gate
node dist/xsquared.js list [--source <source-id>] [--json]
node dist/xsquared.js post <post-id>
node dist/xsquared.js profile-learn --handle "@therealtongchen"
node dist/xsquared.js dashboard --port 3888

Env vars for the LLM pipeline:

• XSQUARED_CLAUDE_MODEL (default sonnet) — model alias passed to claude -p.
• XSQUARED_RESEARCH_BUDGET_USD (default 0.50) — max budget per research call.
• XSQUARED_GENERATE_BUDGET_USD (default 0.50) — max budget per generation call.
• XSQUARED_DISABLE_LLM=1 — force the template fallback (offline / cost-free).
• XSQUARED_CLAUDE_BIN — override the claude binary path.

The Posts tab groups drafts by Topic and Trending. It supports editing drafts, recording rewrite requests for OpenClaw, inspecting what xsquared has learned about your writing style, and posting approved drafts through birdclaw compose post.

Hourly automation is installed as OpenClaw cron `e4dff86f-e451-47cc-bfd6-2cc9b5f033ee` (`xsquared hourly viral auto-drafts`). It runs `npm run hourly-drafts`, creates local drafts only, says `NO_REPLY` when no new qualifying posts are found, and reports in the Telegram x-growth topic only when drafts are created or the job fails.

Viral drafts pass through a deterministic grounding validator before storage. High-risk claims like `can now`, launch claims, `ready-to-post`, workflow automation claims, and vague product promises are rejected unless the selected source post or source context supports them. Regression check: `npm run validate:grounding`.

Auto-drafting has stricter production gates than dashboard feed browsing:

• A viral source must have either a search filter or a relevance focus.
• DeepSeek ranking must succeed before drafts are created.
• Local heuristic ranking is allowed for displaying/debugging feed candidates, but auto-draft will skip it unless `XSQUARED_ALLOW_LOCAL_RANK_AUTODRAFT=1` is explicitly set.

The sidebar includes an Eigen chat panel. It auto-detects the latest OpenClaw Telegram topic session and sends messages back into that session with openclaw agent --session-id; the UI stores its local transcript in .xsquared/store.json.

The automatic content loop is intentionally simple for the first version:

1. Use Topic when you already know what you want to post about, such as "Google Ads for small business".
2. Use Trending when you want xsquared to pull relevant feed posts and help recreate your version of what is working.
3. xsquared combines the selected source with the latest writing-profile snapshot. Topic generation uses learned tweet samples for voice by default unless that source is explicitly turned off.
4. It saves draft posts locally for review and rewrite.

Profile learning runs automatically when the dashboard loads profile data. xsquared first checks Birdclaw's local authored-tweet store, then falls back to Bird's profile timeline fetch for @therealtongchen when local authored tweets are empty.

Persistence: xsquared stores generated drafts, topics, source snapshots, rewrite requests, and learned profile snapshots locally in .xsquared/store.json by default. Set XSQUARED_HOME to move that app-data directory. Bird/Birdclaw source history is read from the local Birdclaw/Bird setup, including ~/.birdclaw.

Privacy: xsquared does not send data to X unless a post action is explicitly triggered.

Terms: Use at your own risk. Review content before publishing.