docs-mcp

Drop any Word, Excel, PDF or PowerPoint into a vector RAG store. Vision-model extraction handles scans, charts, and tables. Every returned chunk carries its page number for precise citations. Priced at exact break-even via Stripe — we make $0 per credit sold.

Live: docs.regiq.in

What it does

1. Ingest — accepts .pdf, .docx, .xlsx, .pptx, .doc, .xls, .ppt up to 50 MB.
2. Extract — renders every page as an image, runs Google Gemini 2.5 Flash Lite (via OpenRouter) over each page. Text, tables, and figure descriptions come out verbatim.
3. Store — chunks (~500 tokens, 50-token overlap), embeds via openai/text-embedding-3-small (1536 dims), stored in Postgres + pgvector with page-number metadata.
4. Retrieve — cosine-similarity search returns the top-k chunks; your agent synthesizes the answer.

Tools

Tool	What it does
docs_upload({filename, contentBase64, mimeType?})	Small (≤~7 MB) programmatic upload. Ingest runs async.
docs_list()	All documents owned by the calling key.
docs_get({id})	One doc's metadata + status + chunk count.
docs_search({query, k?, documentIds?})	Semantic search. Returns top-k chunks with page numbers. Free.
docs_delete({id})	Permanent delete.
docs_balance()	Credit balance + last 10 transactions.

Bigger files: upload via the web dashboard at docs.regiq.in/dashboard (max 50 MB).

Pricing

1 credit = 1 page ingested. Queries are free. New accounts get 100 pages free on sign-up.

Top-up	Pages you get	~Docs (10-pg avg)
$5	5,700	570
$10	11,700	1,170
$20	23,900	2,390
$50	60,500	6,050

Per-page underlying cost is roughly $0.0005 vision + $0.00003 embedding via OpenRouter. Prices are set at exact break-even after Stripe's 2.9% + $0.30 fee — that flat fee is why $2 top-ups aren't offered (18% of $2 evaporates to Stripe).

Setup — any MCP client

1. Sign in at docs.regiq.in with Google or GitHub.
2. Copy your API key from /dashboard.
3. Add to your client config:

{
  "mcpServers": {
    "docs": {
      "url": "https://docs.regiq.in/api/mcp",
      "headers": {
        "Authorization": "Bearer docs_live_..."
      }
    }
  }
}

Works in Claude Desktop, Cursor, Zed, and anything else that speaks streamable-http MCP.

Recommended flow

1. docs_upload({filename, contentBase64})   →  { id, status: "processing" }
2. poll docs_get({id}) until status="ready" (~10-60s for 10 pages)
3. docs_search({query: "...", k: 8})        →  chunks with page numbers
4. Your LLM synthesizes an answer and cites the page numbers.

Self-host

git clone https://github.com/globalion/docs-mcp
cd docs-mcp
cp .env.example .env  # fill in Google OAuth + OpenRouter + (optional) Stripe
docker compose up -d --build

Uses pgvector/pgvector:pg16 image so the vector extension is pre-installed. docker exec docs-mcp-web npx prisma@6.19.2 db push --accept-data-loss --skip-generate on first boot to sync the schema (the container CMD does this automatically).

License

MIT — see LICENSE. Built by Shreyas, shipped by Globalion.