Advice about posting is usually taste; this one quotes the signal weights out of the open-source ranking code and scores against them, which makes a claim you can check rather than agree with. The numbers reward what the code rewards — a reply is worth 27.0× while a report is -369× — so read the score as a fit to those weights, not as a prediction of reach.
A toolkit built on X's open-source recommendation algorithm, shipped as a React web app and an MCP server. The MCP half puts the scoring engine behind eleven tools, so an assistant can score a draft, test its opening line, check it against the quality filters and plan posting cadence.
- A draft scored against the algorithm's signal weights — `score_tweet`, and `full_audit` for everything at once
- The opening line tested on its own, against the penalty that fires when a reader scrolls past — `analyze_hook`
- The 18 quality and safety gates run over a draft — `check_filters`, with `detect_spam_patterns` alongside
- Posting cadence from the author-diversity decay formula — `diversity_penalty` and `get_optimal_schedule`
- Two drafts compared, a niche classified, and thread strategy — `compare_tweets`, `classify_niche`, `get_thread_strategy`
- The raw reference data as MCP resources — `x-algorithm://weights`, `x-algorithm://filters`, `x-algorithm://niches`
- The weights themselves, named and signed: Reply at 27.0×, Bookmark at 10.0×, Follow at 4.0×, against Report at -369× and Not Interested at -74×
For the MCP server, Node. The install lines and the documented Claude Desktop entry both name `x-algorithm-mcp`, which npm does not have. No key is needed for scoring, filters and analysis. The web app is separate: clone, `npm install`, `npm run dev`, and it opens on `http://localhost:5173`; its AI rewrite, generate and thread features are the only part that takes an API key, added in Settings.