Parse any document into structured blocks — and generate documents in 9 formats — using the AILANG Parse API.
Inside Claude Code, add the marketplace and install the plugin:
/plugin marketplace add sunholo-data/docparse-skill
/plugin install ailang-parse@ailang-parse-marketplace
Codex discovers user-wide skills in ~/.agents/skills and follows symlinked
skill folders. Keep this repo as the single source of truth instead of copying
SKILL.md into each agent's configuration. From a clone of this repo:
git pull --ff-only
DOCPARSE_SKILL_REPO="$(pwd -P)"
mkdir -p ~/.agents/skills
ln -s "$DOCPARSE_SKILL_REPO/plugins/ailang-parse/skills/ailang-parse" \
~/.agents/skills/ailang-parseDo not overwrite an existing skill directory or symlink: inspect its target
first. For Claude, either keep the marketplace plugin installation above or
link this same folder into ~/.claude/skills/ailang-parse:
mkdir -p ~/.claude/skills
ln -s "$DOCPARSE_SKILL_REPO/plugins/ailang-parse/skills/ailang-parse" \
~/.claude/skills/ailang-parseAvoid installing both the standalone Claude skill and the plugin's copy if that
would create duplicate skill entries. Linked installations share edits
immediately; a cached Claude marketplace plugin needs its own update. Subsequent
git pull --ff-only updates the linked skill; commit or otherwise preserve local
edits before syncing if they overlap upstream changes.
In Codex, invoke $ailang-parse or ask naturally about docparse, Word document
creation, or conversion. Automatic invocation remains enabled. Codex detects
skill updates; if the skill is absent from the selector, restart Codex. This is
a laptop-wide installation for local sessions using this home directory, not
an installation on remote hosts or cloud environments.
A skill symlink does not install an MCP server. The local CLI works without one. For optional hosted access, register the connection separately:
codex mcp add ailang-parse --url https://docparse.ailang.sunholo.com/mcp/Use the service's device-auth tools if its API requests a key. Installing a skill does not copy Claude credentials or hooks into Codex. See Codex skills and MCP configuration.
This plugin registers an MCP server against the hosted AILANG Parse API, so documents parsed through it are uploaded to the cloud service. That is the right default for general use and needs no setup.
For material that must not leave the machine — confidential, client-privileged,
under NDA — use the local docparse CLI instead. It runs the same parsers
and generators from the public
ailang-parse repo, on your own
hardware:
curl -fsSL https://www.sunholo.com/ailang-parse/install.sh | sh
docparse report.docx --output-dir ./parsedOne command, no clone (ailang_parse 0.40.0+). It installs the AILANG runtime
if you do not have it, fetches the published package (~400 KB, not the 24 MB
repo archive), and puts docparse on your PATH. --version, --prefix and
--uninstall are supported; re-running is a no-op.
That is everything for Office, ODF, HTML, Markdown, CSV, TeX, EPUB and email. PDF needs two more steps that are easy to miss:
brew install poppler # pdftotext, the default backend
# (apt install poppler-utils on Debian/Ubuntu)
docparse --install-backends # docling + liteparse, for scans and layoutWithout docling, a scanned PDF fails even on the default backend —
pdftotext escalates to it automatically when it finds no text layer.
Contributors can still git clone the repo and symlink bin/docparse; that
remains the development workflow.
The two coexist. Once the plugin is installed, the skill carries the decision rule and tells you which path a given parse is using, so you are never uploading a document by accident.
The local CLI also does things the hosted API cannot: files over 32MB, audio and
video, the slow local PDF backends (docling, liteparse — the hosted API is
capped at 30s), and AI generation from a prompt (--generate). Full reference:
plugins/ailang-parse/skills/ailang-parse/resources/local-cli.md.
Caveat: "local" is a property of the backend, not the CLI. Office, ODF,
text, HTML, EPUB, email and pdftotext/docling/liteparse PDFs never touch
the network — but --pdf-backend ai, --describe, --summarize, images and
audio/video all send content to an AI provider.
This plugin registers an MCP server and a skill for document parsing and generation. When installed, Claude automatically gets these tools:
| Tool | Purpose |
|---|---|
mcpParse |
Parse any document into blocks, Markdown, HTML, or A2UI |
mcpConvert |
Generate a document — docx, pptx, xlsx, odt, odp, ods, html, md, qmd |
editDocument |
Parse a document, apply JSON edit deltas, return modified blocks (Office formats only) |
getUploadUrl |
Pre-authenticated GCS upload URL for large files (Business tier) |
mcpFormats |
Discover formats, samples, pricing, capabilities |
mcpEstimate |
Predict cost/latency before parsing |
mcpAuth |
Get an API key via device auth (RFC 8628) |
mcpAuthPoll |
Poll for auth completion |
mcpAccount |
View status, keys/usage, or pricing (no auth for pricing) |
submit_feedback |
Report a bug / feature / docs gap to the maintainers |
Just ask Claude to parse or produce a document — the MCP tools handle everything automatically, including authentication.
# Reading documents:
"Parse this DOCX file and show me the headings"
"Extract tables from report.xlsx"
"Convert this PDF to markdown"
# Producing documents:
"Turn these notes into a PowerPoint"
"Write this up as a Word document"
"Give me that summary as an .xlsx"
"Export this analysis as Quarto"
# About the service:
"What formats do you support?"
"How much would it cost to parse this file?"
To generate a document, Claude writes Markdown and converts it — front matter, tables, inline formatting, links and embedded images all carry through. Headers, footers, comments and tracked changes have no Markdown syntax, so those survive only when converting from a document that already has them.
No manual API key setup needed. If auth is required, Claude will walk you through the device auth flow.
# Set your API key if you already have one
export DOCPARSE_API_KEY="dp_your_key_here"Or get one via:
- AILANG Parse Dashboard — sign in with Google
- Device auth flow — Claude handles this automatically via
mcpAuth
Parsing:
| Category | Formats | Speed |
|---|---|---|
| Office | DOCX, PPTX, XLSX, ODT, ODP, ODS | 5-50ms deterministic |
| Text | CSV, Markdown, HTML, EPUB, EML, MBOX, TEX, RTF | 5-15ms deterministic |
| PDF/Image | PDF, PNG, JPG | AI-powered |
| Audio/Video | WAV, MP3, MP4, and other media | Local CLI only — the hosted API rejects these |
Generation (9): DOCX, PPTX, XLSX, ODT, ODP, ODS, HTML, Markdown, QMD (Quarto)
Ask mcpFormats for the live list — it is the service's own answer and never goes stale.
Per-document pricing (not per-page). A 1-page or 1,000-page document costs the same.
| Tier | Monthly | Requests | AI Parses |
|---|---|---|---|
| Free | EUR 0 | 1,000 | 50 |
| Pro | EUR 29 | 100,000 | 500 |
| Business | EUR 99 | 500,000 | 2,000 |
pi does not load MCP servers, so pi users install the sibling extension
instead: pi-extension/ registers native pi tools
(docparse_parse, docparse_convert, docparse_generate, docparse_status)
plus a /docparse-login device-flow command, calling the local CLI and the
same hosted API (parse, convert and the local-only prompt generation). Install:
ln -s ~/dev/sunholo/docparse-skill/pi-extension ~/.pi/agent/extensions/docparseThe skill includes on-demand writing-quality and template-inspection guidance.
Its local scripts/audit.sh and scripts/render.sh wrappers call the AILANG Parse
companions docparse-audit and docparse-render. These companions are currently
in the Parse checkout's unreleased changes; older published installs may not
include them. Update the engine when released, or set DOCPARSE_AUDIT_BIN and
DOCPARSE_RENDER_BIN to the checkout executables. Rendering/comparison and audits
are implemented in AILANG; Office rendering additionally uses LibreOffice and
Poppler. No new Python dependency is required.