Skip to main content
Glama

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

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)

  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):

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                      # 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.

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.

Available Tools

8 tools
benchmarkB

Run every available backend on the same input, compare results.

Diagnostic tool. Returns one ParseResult per available backend with latency, byte counts, error state, and metadata so you can see at a glance which parser handles a given document best.

Inputs: source (required). Returns: object

chunk_textA

Chunk a markdown document into retrieval-ready pieces.

Picks a doc-type-aware chunker (paper / book / manual / qa / resume / table / default) based on document structure. Set doc_type to force a specific chunker; default "auto" runs structural detection.

Inputs: text (required), doc_type, max_tokens, min_tokens, +1 more. Returns: object

detect_doc_typeA

Run structural heuristics to detect document type.

Useful when you want to know what chunk_text would pick before actually chunking. Returns the resolved doc_type token.

Inputs: text (required). Returns: object

interpretA

Parse the document, then ask Claude over the parsed markdown.

Use this for "summarize this PDF", "extract every action item", "what does this contract say about termination" style requests. The router parses first (cheap, deterministic), then Claude reads the markdown and answers. Cache hits across calls in the same session reuse the parsed text for free input tokens.

Inputs: source (required), instruction (required), model, backend, +1 more. Returns: object

list_backendsA

Report which parse backends are installed + which are missing.

Returns a list of {name, available, module} entries. Use this to debug "why did the router fall back to markitdown" by checking whether docling/llamaparse are actually available.

Inputs: no named parameters. Returns: object

parseA

Convert a document to markdown. Routes across backends automatically.

Inputs: source (required), hints, backend. Returns: object

parse_to_vaultA

Parse a document and write it to the vault as a markdown note.

Replaces the standalone ⚙️ Meta/scripts/markitdown_to_vault.py shell script: same output shape, but with the router's full audit trail (which backend ran, fallback chain, latency) baked into the frontmatter.

Inputs: source (required), backend, overwrite, vault_folder. Returns: object

parse_urlC

Shortcut for parsing an HTTP(S) URL. Same return shape as parse.

Inputs: url (required), backend. Returns: object

Tool Schema Changelog

Recent tool additions, removals, and schema changes observed during successful MCP inspections.

Loading tool schema changelog…

    TDQS

    B3.4/5.0

    Scored across 8 tools

    Disambiguation3/5

    parse, parse_url, parse_to_vault, and interpret all build on the same underlying parse operation, so an agent could easily select the wrong variant when the target or output destination is not explicitly clear. The descriptions help, but the boundaries between these tools remain somewhat fuzzy.

    Naming Consistency4/5

    Most tools follow a predictable snake_case verb_noun pattern such as list_backends, chunk_text, and detect_doc_type. parse_to_vault and single-word verbs like parse, interpret, and benchmark are minor deviations, but the overall naming style is coherent.

    Tool Count5/5

    Eight tools is well-scoped for a document parsing server with diagnostics, vault integration, and chunking support. Each tool maps to a distinct workflow need, and none feel redundant or like filler.

    Completeness4/5

    The toolset covers the full parse-chunk-interpret workflow, plus backend diagnostics and vault output. The main gap is the lack of a batch or multi-document parsing operation, but agents can work around that with repeated calls.

    Maintenance

    ActivityMaintained
    ResponsivenessNo recent community issues

    Related MCP Connectors

    Related MCP Servers