Agent Economy ReportAgent Economy ReportCAUTION version 1.0.0 · audited 2026-09-09
This skill is documentation only: it tells the agent how to run Microsoft's `markitdown` tool via `uvx` to convert PDFs, Office files, images, audio, archives and YouTube URLs into Markdown. There is no script, no installer and no code that runs on its own — just a SKILL.md with example shell commands and a skill card. The only outside network reference is a placeholder Azure Document Intelligence endpoint (`your-resource.cognitiveservices.azure.com`) shown as an optional flag for better PDF extraction, and a publisher profile link on clawhub.ai. It does not read credentials, set up persistence, or send data anywhere hidden. The main things to be aware of are that following the instructions causes `uvx` to download and run the third-party `markitdown` package from the internet, and that the documented `--use-plugins` flag would enable arbitrary third-party plugin code.
| Declared purpose | Convert documents and files (PDF, Office, HTML, CSV, images, audio, ZIP, YouTube URLs, EPub) to Markdown for LLM processing or text analysis, using the markitdown CLI. |
| Observed behavior | Provides static Markdown guidance with bash examples invoking `uvx markitdown` on local files and stdin, plus a list of CLI options. No executable files, no environment variable access, no file writes by the skill itself, no obfuscation, no persistence, no credential or payment handling. Behavior matches the declared purpose. |
| Verdict, rules only | SAFE |
| Verdict, AI | CAUTION |
| Final verdict | CAUTION (never better than either pass) |
| capability | vs purpose | detail |
|---|---|---|
| shell | needed | SKILL.md instructs the agent to run bash commands such as `uvx markitdown input.pdf` (line 14) to perform conversions. |
| filesystem_read | needed | Input documents are read from disk by the recommended command, e.g. `uvx markitdown report.docx -o report.md` (line 48). |
| filesystem_write | needed | Converted output is written to a user-chosen path via `-o output.md` or shell redirection (lines 17-18). |
| install_packages | needed | `uvx` fetches and executes the markitdown package from a package index on first run; SKILL.md line 66 notes 'First run caches dependencies'. |
| network | needed | Network use is limited to package fetch by uvx, optional Azure Document Intelligence endpoint (line 60, a placeholder), and YouTube URL conversion (line 29). No hidden or hardcoded destinations. |
| severity | finding |
|---|---|
| low | Instructs the agent to download and execute a third-party package at runtime `uvx` resolves and runs the `markitdown` package from a remote package index each time it is not cached. This is the standard, documented use of the tool and matches the skill's purpose, but it means unpinned third-party code is fetched and executed on the host. Users in locked-down environments should pre-install and pin a version instead. SKILL.md:14 uvx markitdown input.pdf |
| low | Documents a flag that enables arbitrary third-party plugin code The skill lists the markitdown `--use-plugins` option without a caveat in SKILL.md; enabling it lets any installed third-party plugin run during conversion. The skill card (line 29) does advise enabling plugins only for approved providers. No plugin is bundled or installed by this package. SKILL.md:40 --use-plugins # Enable 3rd-party plugins |
| info | Optional cloud upload of document contents The documented `-d`/`-e` flags send the document to an Azure Document Intelligence endpoint for OCR. The host shown is an obvious placeholder ('your-resource'), not an attacker-controlled destination, and the feature is opt-in, but users should note that using it transmits file contents to a cloud service. SKILL.md:60 uvx markitdown scan.pdf -d -e "https://your-resource.cognitiveservices.azure.com/" |
| info | Converted content is untrusted input The skill's whole function is to turn arbitrary documents, archives and web/YouTube content into text that the agent will then read. Any instructions embedded in those source documents could act as indirect prompt injection against the agent. This is inherent to the conversion task, not a defect introduced by this package. SKILL.md:29 - **Other**: ZIP (iterates contents), YouTube URLs, EPub |
0 AI finding(s) were dropped because their file, line or quote did not match the package.
| domain | service rating | where |
|---|---|---|
| clawhub.ai | not in the service index | skill-card.md:7 |
| your-resource.cognitiveservices.azure.com | not in the service index | SKILL.md:60 |
| Binaries invoked | bash |
| Environment variables read | none |
| Hard-coded wallets | none |
| Pipes a download to a shell | 0 |
| eval / exec / subprocess | 0 |
| base64 blobs | 0 |
| File writes | 0 |
| Persistence | 0 |
| file | lines | sha256 |
|---|---|---|
SKILL.md | 67 | 354b8335051b9037… |
_meta.json | 6 | f6b0d4ff560826a5… |
skill-card.md | 35 | 4ac4aa26d1c25f97… |
For agents
JSON: https://agenteconomy.report/k/markdown-converter.audit.json · badge: https://agenteconomy.report/k/markdown-converter.audited.svg ·
skill rating: /k/markdown-converter · commission an audit of any skill: US$ 29 per version.
The complete published package of this exact version was downloaded from the registry and read statically; nothing was executed. A deterministic pass extracts network destinations, binaries, environment variables, writes, install commands, obfuscation markers, subprocesses and wallets, each with file and line. An AI then reads the whole package with those facts and writes the summary, the capabilities and the findings under a strict schema; every finding it produces must cite a file, a line and the exact text, or it is dropped. The final verdict is the worst of the two passes. The audit does not change the skill's trust tier (policy); the author may respond through the dispute channel and the response is published here. Commissioned by: the Agent Economy Report (free program: the 150 most downloaded skills, September 2026).