Show HN: Open-source multi-agent subtitle translator (self-hosted)

I spent the past year solving a real problem: AI-translated subtitles are inconsistent—switching between formal and casual, character names changing mid-film, tone jumping around randomly.

Existing tools treat subtitles as isolated sentences. So I built a *multi-agent system* that works like a real translation team: a "director" creates a style guide for the entire film, a "glossary agent" enforces consistent terminology, and a "translator" works with context windows.

It works well. I've processed hundreds of hours of subtitles with it, and the quality beats most SaaS products I've tried.

Honestly, I originally wanted to build a SaaS. But getting a product noticed is hard—marketing, customer acquisition, support. As a solo developer, I don't have the resources of a big company or VC.

Rather than let this project die on my hard drive, I'm open-sourcing it so people who need it can deploy it themselves. At least this way, the time I spent isn't wasted, and it can help some people.

Three specialized agents work in sequence: 1. Director Agent - Reads the entire subtitle file and generates a style guide (tone, formality, addressing strategy). Uses long-context models to analyze the full script.

2. Glossary Agent - Extracts and enforces consistent terminology: character names (with gender inference for pronouns), locations, domain terms. Outputs a Markdown glossary embeddable in prompts.

3. Translator Agent - Uses a sliding window with bidirectional context: - References previous translations (for coherence) - Previews upcoming sentences (to avoid splitting errors) - Strict format validation (preserves timing, IDs, structure) - Batch processing for efficiency

4-6. Reviewer, Polisher, and Timing Adjuster agents are planned for Q1 2026.

Stack: - .NET 10.0 + Microsoft Agents framework - Subtitle Edit (libse) for parsing - Multiple subtitle formats (auto-detection) - Compact format to reduce token usage - Configurable LLM endpoints (OpenAI-compatible, Gemini, etc.) - Model strategy: Long-context for Director/Glossary, cost-effective for translation

Status: [Y]Core pipeline (steps 1-3) is production-ready and open source [Y]Monolingual translation works [Y]Bilingual subtitle support available Steps 4-6 Before Q2 2026 Windows/macOS/Web UI planned for May 2026