About
Falsafa
A reading platform for translated philosophical and classical works, plus an open-source MCP engine that lets any LLM query the corpus as a knowledge resource.
What's in the catalog
37 works, 818 logical chapters, 2,053 language variants. The corpus spans Anglo-Saxon Christian poetry (Cynewulf), Urdu ghazal masters (Ghalib, Iqbal, Zauq), French Enlightenment political theory (Comte, Dunoyer), German philosophical writing (Fichte), Sanskrit smṛti traditions, and Old Javanese / Kawi tattva texts. Most works ship with 2-3 variants — original-language source, Latin-script transliteration, and English translation. Live counts at /numbers.
Sources
The source texts the corpus is built from come from freely available digital archives plus carefully transcribed printed editions. We acknowledge each archive directly:
- Old English works (Cynewulf's Juliana, Elene, Andreas; the Old English Elegies) — drawn from sacred-texts.com, a long-standing public-domain archive of religious and classical literature.
- Indian texts (the Sanskrit smṛti corpus — Manusmṛti, Yājñavalkya, Viṣṇu, Nārada, Bṛhaspati, Parāśara, Aṅgirasa, and others) — drawn from GRETIL, the Göttingen Register of Electronic Texts in Indian Languages.
- Allama Iqbal's poetry (the three parts of Bāng-i-Darā) — drawn from allamaiqbal.com, the Iqbal Academy Pakistan archive.
- Mirza Ghalib and Sheikh Ibrahim Zauq (the two Urdu diwans) — transcribed from various printed editions.
Where the source is freely available online and we drew from it, the link above goes directly to the archive. Where we worked from printed books, we acknowledge that here. The translations and transliterations layered on top of these sources are described in the next section.
Thothica
The translations and transliterations were produced by Thothica, an AI translation company. Thothica's pipeline draws on frontier large language models — Anthropic's Claude, OpenAI's GPT, Google's Gemini, and open-weights models via OpenRouter. Without those models, the catalog at this scale would not exist.
Every chapter page surfaces the credit verb that matches what was done — translated, transliterated, or curated.
The MCP engine
Falsafa ships an open-source MCP server (@falsafa/mcp) that
exposes the corpus as eight librarian tools — list works, list
chapters, read a chapter, get a passage, search the corpus, find
related, compare works, and read metadata. Plug it into Claude Desktop,
Cursor, Codex, or any MCP-aware client, and your chosen LLM can
navigate the catalog directly.
The MCP is a librarian, not a second LLM. Tools return text and structure; the host LLM does the reasoning. No API keys, no inference cost, no model lock-in. Inspired by Karpathy's vector-DB-less RAG gist .
The covers
Every cover is a watercolor painting produced by an agentic pipeline
(Claude Sonnet 4.6 drafts → GPT-5.5 critiques → Claude Sonnet 4.6
decides → gpt-5.4-image-2 renders). Cross-vendor critique catches
AI-slop, hedging, and palette drift before the prompt reaches the
image model. Every stage's I/O is persisted to a per-work
cover.audit.json for full reproducibility.
Series with multiple works (Cynewulf trilogy, Iqbal's three Bang-E-Dara parts, Comte's volumes, the Sanskrit smṛti corpus, the Kawi tattva corpus) share a palette + mood anchor so siblings read as related; works in different series look genuinely different.