Document Tools plugin

mere-doc-tools converts office documents and text PDFs to Markdown with AnyDoc, or performs image OCR with native mere.run models. Both paths support optional local PII redaction. The command is installed by the mere-workflow-tools package.

Install

mere.run plugin install mere-doc-tools --yes
mere-doc-tools doctor --extractor anydoc --no-redact

Python compatibility

On Python 3.10+, the normal installation includes AnyDoc automatically. No extra or API key is needed. Catalog installation also registers the document graph provider with mere.run. Python 3.9 installations retain native OCR and the other workflow tools, but document conversion requires a newer interpreter.

To select a specific interpreter when installing directly with pipx:

pipx install --python python3.12 "git+https://github.com/sawfwair/mere-run-plugins.git@main#subdirectory=packages/mere-workflow-tools"
mere-doc-tools doctor --extractor anydoc --no-redact

Choose an available Python 3.10+ interpreter for --python. Upgrade an existing installation with pipx upgrade mere-workflow-tools. To migrate a Python 3.9 environment, use pipx reinstall --python python3.12 mere-workflow-tools. The former [anydoc] extra remains a compatibility alias. For an older plugin version already running on Python 3.10+, the dependency can also be added with:

pipx inject mere-workflow-tools "firecrawl-anydoc>=0.2.4,<0.3"

The distribution is firecrawl-anydoc, imported as anydoc; the unrelated package named anydoc on PyPI is not supported. doctor --extractor anydoc checks both conversion and the default redaction dependencies. Adding --no-redact checks conversion without requiring a mere.run executable.

Convert a document to Markdown

mere-doc-tools process \
  --extractor anydoc \
  --input ./report.docx \
  --output-dir ./doc-out \
  --no-redact

This writes report.md and a durable run.json, including the executed AnyDoc version and SHA-256 hashes of output artifacts. No model or mere.run executable is needed when redaction is disabled. AnyDoc runs locally in Python through its Rust bindings; the plugin never invokes npx, downloads a runtime on demand, or enables hosted OCR.

AnyDoc supports Word (doc, docx, docm), PowerPoint (ppt, pps, pot, pptx, pptm, ppsx, ppsm), Excel (xls, xlsx, xlsm, xlsb), OpenDocument (odt, ods, odp), RTF, EPUB, CSV, and text-based PDF. It detects formats from content, with the file extension as fallback, including for CSV. This workflow converts one local file per run.

Omit --no-redact to also write report.redacted.md and report.pii.json using native mere.run text anonymize. The original unredacted Markdown remains in the output directory. Review redaction results before sharing; conversion does not preserve page layout or export embedded binary assets.

OCR a document image

mere-doc-tools process \
  --input ./scan.png \
  --output-dir ./doc-out \
  --ocr-backend lighton \
  --redact

--extractor ocr is the default, preserving existing image workflows. Redaction is enabled by default for both extractors. Use --no-redact to skip it, or --replacement to change the replacement format. Supported native OCR backends are lighton, glm, and infinity; --ocr-backend only applies to the ocr extractor.

Plan and run separately

mere-doc-tools plan \
  --extractor anydoc \
  --input ./report.docx \
  --output-dir ./doc-out \
  --run-id document-001 \
  --no-redact
mere-doc-tools run ./doc-out/run.json
mere-doc-tools resume ./doc-out/run.json
mere-doc-tools cleanup ./doc-out/run.json

plan and process --dry-run write the plan without loading AnyDoc or running models. run --dry-run prints the existing plan without execution. run executes it, including when retrying a failed run after installing a missing dependency. resume inspects recorded state; it does not rerun conversion. Cleanup records a local no-op and retains source and derived files.

An executable synthetic example is in examples/documents.

Native workflow graphs

The plugin advertises graph-node-provider-v1 and the document.convert node. The node accepts one local input asset and emits:

Output	Type	Content
markdown	asset	UTF-8 Markdown file
manifest	asset	Durable plugin-run.v1 receipt, compatible with run, resume, and cleanup
text	string	The same unredacted Markdown, usable by downstream text nodes
stats	JSON	Source/output SHA-256, output byte count, and AnyDoc version

mere-doc-tools graph catalog --json
mere-doc-tools graph templates export document-to-markdown --output ./document.graph.json
printf '{"document":"./report.docx"}\n' > document.inputs.json
mere.run graph preflight ./document.graph.json --inputs-json ./document.inputs.json --json
mere.run graph run ./document.graph.json --inputs-json ./document.inputs.json --run-dir ./document-graph-run --json

For direct pipx installations that have not been registered by the catalog installer, set MERERUN_GRAPH_PROVIDERS=mere-doc-tools for the native graph commands. Install the plugin on each executor that runs this node. Graph execution uses no model, credentials, or network access; scanned PDFs fail locally. Graph outputs are not redacted. Use the CLI redaction workflow before sharing sensitive text.

Preflight checks the backend, input, and confined output paths without converting or creating a run directory. Execution writes a running manifest before conversion, persists failures, and emits NDJSON progress, artifacts, statistics, and one final result only after success. It refuses source/output aliases and detects source changes during conversion. Cross-run caching is disabled because the AnyDoc version can change independently of the plugin version.

Conversion failures

Missing or incompatible AnyDoc installations return exit code 3 with install guidance. Conversion failures return 1 and persist a failed run manifest. Encrypted, malformed, unsupported, or resource-limited documents do not silently succeed. Scanned or image-only PDF pages fail locally with an OCR-required diagnostic. Render those pages to images and use --extractor ocr; there is no automatic fallback or document upload.

Privacy boundary

No hosted API is used. AnyDoc is always called with ocr="reject", even when FIRECRAWL_API_KEY or FIRECRAWL_API_URL is set. Source and derived text stay local, but callers remain responsible for access control and review of redaction quality. See Private workflows.

Third-party notices

The package includes mere_workflow_tools/THIRD_PARTY_NOTICES.txt in its source distribution, wheel, and installed files. It preserves the full MIT notices for AnyDoc 0.2.4 and its pdf-inspector 1.14.2 component.

If you bundle or redistribute dependencies, preserve their applicable license and copyright notices. This file supplements those notices; it isn't a complete inventory of transitive dependency licenses. Review the resolved dependency versions before redistribution.