parse-mcp

One MCP, many parsers. Default markitdown (free, fast, MIT). Escalate to Docling (table-heavy, scanned PDFs) or LlamaParse (cloud, BYOK) when markitdown's quality isn't enough. Plus an interpret tool that pipes parsed markdown into Claude for "summarize / extract X" so you stop juggling parsers and anthropic skills.

Install

Open Claude Code, paste:

/plugin marketplace add adelaidasofia/parse-mcp
/plugin install parse-mcp@parse-mcp

Manual install (pre-plugin-marketplace). See SETUP.md for full details.

pip3 install --break-system-packages -r requirements.txt
pip3 install --break-system-packages 'markitdown[pdf,docx,pptx,xlsx]'

Then register the server in your client's .mcp.json:

{
  "mcpServers": {
    "parse": {
      "command": "python3",
      "args": ["/absolute/path/to/parse-mcp/server.py"]
    }
  }
}

Related MCP server: Document Parser MCP

Tools

Backends (priority order)

1. markitdown (default, MIT, base install). PDF, DOCX, PPTX, XLSX, HTML, CSV, JSON, XML, EPub, ZIP. Fast, deterministic.

2. docling (optional, pip install docling). Best for complex tables (97.9% on benchmark) + scanned PDFs. Downloads model weights on first run.

3. llamaparse (optional, BYOK, pip install llama-cloud-services + LLAMA_CLOUD_API_KEY). Cloud, cleanest output on visually-complex PDFs.

Routing strategy

• parse(source) with no backend arg: router picks based on file format, falls back if backend errors or returns empty.

• parse(source, backend="docling"): force a specific backend, no fallback. Diagnostic mode.

• Unavailable backends are skipped (logged in the chain), never errored.

Parse-fidelity eval

The routing table above used to be a guess. tests/eval/ turns it into data: a synthetic fixture corpus (16 documents across digital PDF, scanned/image-only PDF, table-heavy, multi-column, and raster image classes) with derived ground-truth markdown, scored against each backend's output on three OmniDocBench / PubTabNet metrics — text edit distance, table TEDS (tree-edit-distance similarity), and reading-order. All scores are quality in [0, 1], higher is better.

Headline result (full table: tests/eval/parse_fidelity_matrix.md):

markitdown is great on clean digital text and digital tables (free, fast, deterministic) but has no OCR — it scores zero on scanned PDFs and images — and it interleaves multi-column layouts. docling wins every class via OCR + layout analysis, at the cost of model-weight downloads. That is the evidence behind the format-preference chain (escalate image/scanned/multi-column to docling first).

Run it:

pip install docling                      # the escalation backend under test
python tests/eval/generate_fixtures.py   # rebuild the corpus (needs fpdf2 + Pillow)
make eval                                # -> parse_fidelity_matrix.{md,json}

The matrix records its provenance (backend + python versions + a fixture-set hash), so a stale result is visible — regenerate with make eval whenever a parse backend is upgraded or retuned. It also reports median latency per backend (the cost axis): the highest-fidelity backend (docling) is far slower than the default, so the router escalates to it rather than defaulting to it.

The scorer's metric tests are pure-Python and backend-free, so pytest tests/ gates them in CI with only the base (markitdown) install — a routing regression that breaks the "markitdown has no OCR" assumption fails the build.

Architecture

FastMCP v3.2.3+, stdio transport, Python 3.13+. Registered in [VAULT_ROOT]/.mcp.json. No daemons, no listeners, no model weights downloaded by default.

See SETUP.md for install + per-backend opt-in.

Related MCPs

Same author, same architecture pattern (FastMCP, draft+confirm on writes, vault auto-export where applicable):

• slack-mcp — multi-workspace Slack

• imessage-mcp — macOS iMessage

• whatsapp-mcp — WhatsApp via whatsmeow

• apollo-mcp — Apollo.io CRM + sequences

• google-workspace-mcp — Gmail / Calendar / Drive / Docs / Sheets

• substack-mcp — Substack writing + analytics

Telemetry

This plugin sends a single anonymous install signal to myceliumai.co the first time it loads in a Claude Code session on a given machine.

What is sent:

• Plugin name (e.g. slack-mcp)

• Plugin version (e.g. 0.1.0)

What is NOT sent:

• No user identifiers, names, emails, tokens, or API keys

• No file paths, message content, or anything from your work

• No IP address is stored after dedup processing

Why: Helps the maintainer know which plugins people actually install, so attention goes to the ones that get used.

Opt out: Set the environment variable MYCELIUM_NO_PING=1 before launching Claude Code. The hook will skip the network call entirely. Already-pinged installs leave a sentinel at ~/.mycelium/onboarded-<plugin> — delete it if you want to reset state.

License

MIT. See LICENSE.

Full install or team version at diazroa.com.