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
Legacy install
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"]
}
}
}
| Tool | What it does |
|---|
parse(source, backend?, hints?) | File path or http(s) URL to markdown. Router picks backend, falls back on empty/error. Returns markdown plus a chain of every backend attempted. |
parse_url(url, backend?) | Shortcut for HTTP(S) inputs. Same return shape as parse. |
parse_to_vault(source, vault_folder?, backend?, overwrite?) | Parse + write the result as a markdown note in the vault. Default folder: <VAULT_ROOT>/📥 Inbox/Converted/. Frontmatter records source, format, backend, latency, bytes_in. Replaces the standalone markitdown_to_vault.py shell script. |
interpret(source, instruction, backend?, model?, max_tokens?) | Parse first, then ask Claude over the parsed markdown. Cache hits reuse parsed text for free input tokens. |
list_backends() | Which backends are installed + which are missing. Diagnostic. |
benchmark(source) | Run every available backend on the same input. Compare latency + output side by side. |
chunk_text(text, doc_type?, target_tokens?, max_tokens?, min_tokens?) | Chunk parsed markdown into retrieval-ready pieces using a doc-type-aware chunker. doc_type="auto" (default) runs structural detection and picks one of paper / book / manual / qa / resume / table / default. Each chunker honors document shape (e.g., paper keeps the abstract whole; manual never merges across numbered sections; qa pairs each question with its answer). Returns chunks + the resolved doc_type. See chunkers/ package. |
detect_doc_type(text) | Diagnostic. Run structural heuristics over markdown and return the doc_type that chunk_text would pick. |
Backends (priority order)
- markitdown (default, MIT, base install). PDF, DOCX, PPTX, XLSX, HTML, CSV, JSON, XML, EPub, ZIP. Fast, deterministic.
- docling (optional,
pip install docling). Best for complex tables (97.9% on benchmark) + scanned PDFs. Downloads model weights on first run.
- 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):
| doc-class | markitdown (text) | docling (text) |
|---|
| digital_pdf | 0.95 | 0.97 |
| table_heavy | 0.93 | 0.89 |
| scanned_pdf | 0.00 | 0.87 |
| image | 0.00 | 0.90 |
| multicolumn | 0.35 | 1.00 |
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
python tests/eval/generate_fixtures.py
make eval
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.
Same author, same architecture pattern (FastMCP, draft+confirm on writes, vault auto-export where applicable):
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.