AI Video Pipeline and MCP Server for Astrology

AI Content Engine for Astrology turns a date into a scheduled avatar Reel and gives LLM agents calculations they can trust.

58 finished Reels produced by the pipeline

~20 min from date and signs to a ready Reel, hands-off

0 code changes to launch a new AI persona

Partner's Profile

Background

The client is an astrology and esoteric media project with three AI personas on Instagram and Telegram. Content follows fixed formats: daily, weekly, 10-day and monthly horoscopes plus affirmation series. Each piece needs a script, a voiced avatar video, a cover, a caption, a first comment and a scheduled post, which does not scale by hand. On Telegram, an LLM-based bot talks to users, and astrology depends on exact planet positions and aspects that a language model cannot be trusted to produce.

Objective

Two goals. First, turn a date and a set of zodiac signs into a finished Reel with cover, caption and first comment, schedule it automatically, and make adding a new persona possible without code changes. Second, give LLM agents a reliable source of astrological calculations, from natal charts and aspects to transits and synastry, so that the model only interprets numbers the server has computed.

Solution

As part of our AI software development work, we built a Python and ffmpeg pipeline that turns a date and zodiac signs into a finished Reel in 5 horoscope formats. An LLM writes the texts, HeyGen generates the avatar video, Gemini creates covers from the persona's face reference, subtitles are built from SRT, B-roll comes from a tagged library of about 40 clips per persona, and ffmpeg mixes the music. Each persona is a config directory with voice, avatar, B-roll and covers, so a new persona means a new config, not new code.

One text cannot serve both speech synthesis and subtitles. The LLM therefore writes two versions: one for TTS, with stress marked in capitals and numbers spelled out, and a normal one for the screen. Publishing moved to the Metricool API: upload via a presigned S3 URL, wait for processing, create a scheduled post. Posts were scheduled across about eight weeks, from 30 April to 22 June 2026, and 49 of 51 publication records received a post ID.

The second part is an MCP server built on one principle: the server calculates, the model interprets. It exposes 10 tools (natal chart, aspects, moon phase, synastry, composite chart, transits, planetary return, geocoding, saving a subject and SVG/PNG chart rendering) plus two interpretation prompts. It runs over stdio for Claude Desktop and Claude Code, and over streamable HTTP with a bearer token for n8n. It was deployed on Google Cloud Run behind a Telegram bot built on n8n, which uses it to answer users and book consultations.

The details decide whether agents can rely on it:

calculations run offline: coordinates and timezone are passed explicitly, geocoding is optional;
coordinates are never rounded, and subjects are cached by a hash of birth data;
chart SVGs of 100–300 KB are saved to a file, and the model gets a link instead of the context-window load;
78 tests, strict mypy and CI on Python 3.11 and 3.12; a reference natal chart matches to 0.001°;
the Swiss Ephemeris core is AGPL, so the service runs internally; SaaS use would need open-sourcing or a commercial ephemeris API.

From Date to Scheduled Reel

Input is a date and zodiac signs; output is a finished Reel with cover, caption and first comment, already scheduled. Daily, weekly, 10-day and monthly horoscopes and affirmation series run through the same pipeline.

One Script for Voice, One for Screen

Speech synthesis and subtitles need different text. The LLM writes a TTS version with stress marked in capitals and numbers spelled out, and a clean version for on-screen subtitles.

Personas as Configuration

Each AI persona is a config directory with its voice, avatar, B-roll library of about 40 tagged clips and cover style. Adding a persona means adding a config, with no changes to the pipeline code.

Server Calculates, Model Interprets

LLMs hallucinate numbers, while charts must be exact. The MCP server computes natal charts, aspects, transits, synastry and more with Swiss Ephemeris, and the agent only explains the results it receives.

Charts Outside the Context Window

Rendered chart SVGs of 100–300 KB are saved to a file, and the model receives only a link. Subjects are cached by a hash of birth data, and coordinates are never rounded.