Indian Liberals MCP Server
adoistic/indianliberals
Acesso via MCP ao arquivo liberal indiano com metadados IA e busca multilíngue.
Ver no GitHubUm servidor MCP que expõe um arquivo digital curado da tradição liberal indiana — 1.225 páginas, 944 obras primárias com metadados extraídos por IA (autores, temas, citações cruzadas, resumos). Permite que LLMs consultem diretamente o corpus com extração de menções entre pensadores, busca multilíngue e citações estruturadas em 4 idiomas.
Instalação
npx convez-ai adoistic/indianliberals1 comando no terminal: detecta o formato do repositório (marketplace ou plugin avulso) e instala via Claude Code. Requer Node 18+ e o Claude Code instalado.
Métricas
Casos de uso
- Pesquisadores consultando corpus de pensadores liberais indianos via agentes de IA
- Indexação e busca semantizada de obras primárias com contexto de autores e temas
- Rastreamento de citações cruzadas e menções entre pensadores históricos
Veredito

Projeto ambicioso e bem-estruturado de arquivo digital acadêmico com camada de IA. A ponte MCP é sofisticada (metadados extraídos, busca multilíngue, cross-references), mas ainda em fase de consolidação — dados parciais (220/944 PDFs processados) e ainda não deployed em produção. Recomendado para pesquisadores e construtores de agentes que entendem o ciclo de vida de corpora em evolução.
README
A tradução para o português deste README ainda está na fila.
Indian Liberals
A modern home for the Indian liberal tradition. This is the in-progress rebuild of indianliberals.in — a digital archive of primary works by historically significant Indian liberal thinkers, alongside thinker profiles, organisation pages, curated excerpts, opinion pieces, interviews, and a federated feed from ThePrint's "Indian Liberals Matter" column.
Built for the Centre for Civil Society by Thothica, in partnership with the Friedrich Naumann Foundation for Freedom.
What this rebuild changes
The legacy site runs on WordPress. Indexing is shallow, metadata is inconsistent, and nothing is machine-readable in the way researchers — human or AI — increasingly work. This rebuild ships:
- A fast, static Astro site on Cloudflare, with a Git-based CMS so editors never see a database.
- A rich AI-extracted metadata layer over the primary-works PDFs — author normalisation, year, themes, publisher provenance, AI summaries, key points, cross-thinker mentions with verbatim evidence — language-agnostic at the schema level.
- An AI/agent layer:
.mdsiblings on every page,/llms.txtand/llms-full.txt, an/AGENTS.mdschema, and (post-launch) an MCP server atmcp.indianliberals.infor frontier LLMs to query the corpus directly. - Multilingual search via Pagefind with per-language analyzers — English alongside Hindi, Gujarati, Marathi, Bengali. 4 languages currently indexed; 1,225 pages, ~28,000 words.
- An honest two-tier model: clean content (musings, opinions, interviews, profiles, organisation pages, the ThePrint mirror) is fully searchable and paragraph-citable; primary-work PDFs surface as rich metadata + AI summary + link, with paragraph-level citation deferred to a future engagement once vision-language layout reconstruction matures.
Where we are
The site builds locally and renders production-quality output. End-to-end content pipeline runs cleanly. Not yet deployed to a public URL — that's Day 10-13 work.
Snapshot (2026-05-18)
| Surface | State |
|---|---|
| Astro site | Builds clean. 1,225 pages, 28,000 indexed words across 4 languages. |
| Content collections | 7 wired: thinkers, organisations, musings, opinions, interviews, primary-works, theprint-mirror. Periodicals empty (deferred). |
| AI extraction pipeline | Production. 220 of ~944 PDFs baked. Rate-limit-aware overnight runner crunches the rest over the coming days. |
| Authority files | 462 thinkers + 50 organisations. byline_lookup at 678 keys. 125 entries flagged bio_source: ai_drafted_stub awaiting real bios (Phase 1.5). |
| Cross-link audit | Phase A applied. 65% of Tier-A entries carry a structured thinker reference (up from 31% at baseline). Bio pages render "By X / About X / Mentioned in" sections with real counts. |
| Cross-thinker mentions | 1,347 mentions across 196 works projected from extraction summaries into primary-work related_thinkers[]. Touchstones like Shroff (72 mentions), Nehru (42), Gandhi (26), Marx (15), Hayek (9), Adam Smith (11) now surface on their bio pages with the works that quote them. |
| TF-IDF related-links | 1,958 cross-collection links computed; rendered as "Related across the archive" on every detail page. |
| Search | Pagefind wired; / keyboard shortcut; type-filter pills; per-language tokenization (Devanagari, Gujarati, etc.). |
| Agent layer | .md siblings on every Tier-A detail page (~580 files). /llms.txt (curated index) and /llms-full.txt (4 MB single-file corpus dump) emit at build time. /AGENTS.md documents the citation rules + tier system. |
| ThePrint federation | RSS-ingest Cloudflare Worker scaffolded + 16/16 vitest passing. List page and homepage feed link out to theprint.in for humans; per-article mirror pages strip the body and keep only metadata + outbound CTA. The body is preserved in the .md sibling for AI agents. NOT yet deployed. |
| Sveltia CMS auth proxy | Worker scaffolded (apps/auth/). NOT yet deployed. |
What remains before launch
These are tracked in docs/superpowers/specs/ and the engagement plan at ~/.gstack/projects/IndianLiberalsWebsite/.
-
Phase B — in-prose NER pass (
docs/superpowers/specs/2026-05-18-phase-b-ner-handoff.md). For every Tier-A entry (musings, opinions, interviews, theprint-mirror), an LLM reads the body and emits structuredthinker_mentions[]records — each with the named thinker, a verbatim evidence quote, and a one-line reasoning. Subject-role entries get 2-4 curated key passages instead of a single quote. The bio page renders a "How X is discussed in this archive" section that aggregates the evidence across the corpus. Pure LLM; no regex. -
Phase 1.5 — real bios for 125 stub thinkers. The Phase A cross-link audit auto-created minimal
ai_drafted_stubentries for every byline that wasn't already in the authority. They render today as placeholder pages. A later pass usesclaude -pto draft 2-4-paragraph bios from the works each thinker authored / pieces about them already in the corpus. -
Day 10-13 deployment work:
- R2 upload of all ~944 PDFs with resumable batch script (PDFs are still on a curator's external drive)
- Cloudflare Pages staging deploy
- Sveltia CMS config + GitHub OAuth app wiring
- ThePrint cron worker deploy + secrets
- DNS cutover prep
-
Day 14 — DNS cutover and launch.
-
3-month support window items captured in
TODOS.mdand the spec docs.
Current limitations (honest)
This is a deliberate list of what is NOT yet good enough. Each is either deferred to a known later pass or scoped out of the engagement.
- Visual / UI design pass. The UI today uses the project's baseline design system (Source Serif, Source Sans, Indic Noto-Serif fallbacks, a saffron + forest accent palette) but no professional design polish. A proper visual pass — typography rhythm, density tuning, image treatment, hero treatments per content kind, mobile interaction details, dark mode — is its own workstream and has not happened yet. The site is functional and readable; it is not yet crafted.
- In-prose Wikipedia-style mention linking. Today the bio pages list works/mentions but the prose inside an article doesn't hyperlink every mentioned thinker's name. Phase B above adds that surface.
- Layout reconstruction of primary works. PDFs are linked but not full-text-indexed. Paragraph-level citation inside the original works is explicitly deferred — it requires reliable vision-language layout reconstruction that clears editorial review, which is a separate engagement.
- Karpathy-style synthesised wiki layer. The "LLM-as-wiki-author" enrichment on top of Tier B is roadmapped post-launch.
- 125 stub thinker bios are minimal one-liners marked
bio_source: ai_drafted_stub. Real bios are a Phase 1.5 task. - The full corpus is not yet extracted. ~220 of ~944 PDFs are baked; the rest are still queued in the overnight runner. Bio pages will grow more works/mentions as the queue drains over the coming days.
- Periodicals collection is empty. Periodical issues live inside primary-works (
work_type: periodical_issue) and are browsable by series at/periodicals/; migrating them into the dedicated collection is deferred. - Editorial review of AI-extracted content is sparse.
needs_review: trueis set on AI-emitted entries to flag this; CCS editorial works through them over time via Sveltia. - MCP server is deferred. The agent layer today is
.mdsiblings +/llms.txt+/AGENTS.md. The MCP server with 8 tools (get_work_metadata,read_clean_content,find_related, etc.) lands post-engagement. - No live deployment. The site builds and previews locally on
:4321; it is not yet served atindianliberals.in.
Stack
Every choice below is already in production on either Falsafa (Thothica) or Liberty Lighthouse (CCS + Thothica). No experimental components.
| Layer | Choice |
|---|---|
| Site generator | Astro 5 with Cloudflare adapter |
| CMS | Sveltia CMS (Decap rewrite, Git-based) |
| Hosting | Cloudflare Pages + Workers + R2 |
| Search | Pagefind with per-language analyzers |
| Metadata extraction | Frontier LLM via headless claude -p CLI, parallelised, rate-limit-aware |
| PDF serving | Direct R2 links (pdf.js viewer not in scope for v1) |
| MCP server | Cloudflare Workers, deferred to post-engagement |
| Repo layout | Monorepo (apps/, scripts/, data/) |
Repository layout
apps/
site/ Astro site
src/
content/ Content collections (7 kinds; 1,200+ entries)
content.config.ts Zod schemas (with shared sub-schemas in src/schemas/)
pages/ Routes incl. per-language /<lang>/<collection>/<slug>
components/ Cards, Search dialog, Related sections, Header, Footer
layouts/ BaseLayout with hreflang + Pagefind body marker
theprint-ingest/ Cloudflare Worker — daily RSS cron mirror with admin-edit guard
auth/ Cloudflare Worker — Sveltia OAuth proxy (scaffolded)
scripts/
llm-extract/ v1.5 extraction pipeline (driver.py, run_overnight.py)
synthesis/ Authority cleanup, TF-IDF, cross-link audit, byline resolution
prompts/ Canonical prompts (system-resolver.txt, README.md);
Phase B will add system-ner.txt here.
data/
authority/ thinkers.json (462), organisations.json (50)
bake-off-output/ Per-PDF extraction outputs (metadata.a.a.json, summary.json)
synthesis/ Aggregations, cross-links, audit residuals, resolution logs
docs/
superpowers/specs/
2026-05-18-cross-link-audit-design.md Phase A spec (applied)
2026-05-18-phase-b-ner-handoff.md Phase B spec (next session)
SCHEMA.md Cataloging schema — the librarian-grade input to extraction +
the editor handbook
ARCHITECTURE.md Top-level design doc
TODOS.md Post-engagement / support-window backlog
Reference projects
- Falsafa — Thothica's reading platform for philosophical and classical texts. 37 works, 818 logical chapters, 2,053 variant entries across 6 languages, 3.1M words. Eight-tool MCP surface, paragraph-stable IDs, deterministic eval at
/eval. The methodological precedent for the no-embeddings, LLM-navigates-markdown approach this site extends. - Liberty Lighthouse — CCS's classical-liberal resource on Indian policy, built with Thothica. Astro plus Decap CMS today; Sveltia is the natural next step. Indian Liberals is the third application of the same architectural template.
- Karpathy's LLM-as-Wiki gist — The methodological source. Humans curate sources, the LLM does the bookkeeping that makes the knowledge base usable, and the wiki is a persistent compounding artifact rather than a chunk index rebuilt on every query.
License
MIT. See LICENSE.
The intellectual content of the archive (primary works, summaries, curated excerpts) is owned by the Centre for Civil Society and the original authors; this repository covers the build infrastructure only.