| name | document-reader |
| description | Load local documents (PDFs, images, handwritten scans, JSON/CSV, text) via the document-reader MCP tool so agents can extract and summarize content without re-implementing file I/O. |
Document Reader Skill
Use this skill when the user provides (or references) local files and you need a reliable way to ingest them into the agent workflow:
- Unstructured docs: PDFs, images, scanned or handwritten notes (returned as base64 + mime)
- Structured docs: JSON, NDJSON, CSV/TSV (returned as text plus optional parsing)
- Plain text: TXT/MD/etc.
This avoids ad-hoc “write a quick Python function to read X” every time.
Prerequisites (Local)
- Start the MCP server:
./scripts/local-test.sh document-reader 7078
- Ensure
.vscode/mcp.json includes:
local-document-reader: http://localhost:7078/mcp
Tool: read_document
When To Use
- The user says “see attached PDF/image” or provides a file path
- You need to extract fields from a form, notes, or report
- You need to load structured input (JSON/CSV) for transformation/validation
Safety Defaults
- Reads are workspace-only by default. To read outside the repo, set
allow_outside_workspace=true.
- Large files are blocked or truncated via
max_bytes / max_chars.
Typical Calls
Read a PDF or image as base64 (for downstream OCR/vision or archive):
{
"path": "data/intake/scanned_note.jpg",
"mode": "binary",
"include_data_url": true
}
Read JSON (returns both text and parsed json when valid):
{
"path": "data/sample_cases/prior_auth_baseline/pa_request.json",
"mode": "text",
"parse_structured": true
}
Read CSV (returns rows up to max_rows):
{
"path": "data/input/patients.csv",
"mode": "text",
"max_rows": 200
}
Suggested Workflow
- Call
read_document for each referenced file path.
- For binaries (PDF/images), use the returned
mime + data_url/base64 to drive downstream extraction.
- Produce a de-identified structured summary (never commit PHI or secrets).
