concinno-skills-office-advanced

Advanced AI-native Office document skills for Concinno. Sibling package to concinno-skills-office.

Status

MVP (0.1.0) — three tools:

⚠ Heavy dependency footprint

docling transitively depends on PyTorch plus vision / OCR model weights. Installing this package adds roughly 2 GB to the environment, and the first PdfAiExtract.call downloads model weights from HuggingFace (cached under ~/.cache/huggingface).

Prefer RunPod / high-spec desktops for benchmarks or heavy batches. A CPU-only laptop can run the pipeline but the first-page latency on a 10-page PDF is typically tens of seconds, not hundreds of milliseconds.

Install

pip install concinno-skills-office-advanced

Requires concinno >= 2.15.1. The MCP bridge adapter lives in Concinno core (concinno.tools.mcp_bridge) — this package depends on it rather than re-implementing JSON-RPC.

Usage via Concinno ToolRegistry

When the consumer sets CONCINNO_LOAD_PLUGINS=1, the default registry auto-mounts all three tools:

import os
os.environ["CONCINNO_LOAD_PLUGINS"] = "1"

from concinno.tools.registry import get_default_registry

reg = get_default_registry()
for name in ("PdfAiExtract", "DoclingPageImage", "AnthropicOfficeMcpBridge"):
    assert name in reg.list_deferred()

Direct Python usage

PdfAiExtract — LLM-ready markdown + tables

from concinno_skills_office_advanced import PdfAiExtract

out = PdfAiExtract().call(
    action="extract",
    path="./report.pdf",
    output_format="markdown",  # or "json" / "text"
)
# out = {
#     "ok": True,
#     "markdown": "# Report …",
#     "tables": [{"page": 1, "html": "<table>…</table>", "csv": "a,b\n1,2\n"}],
#     "page_count": 12,
# }

DoclingPageImage — PDF page → PNG for multimodal LLM

from concinno_skills_office_advanced import DoclingPageImage

out = DoclingPageImage().call(
    action="render",
    path="./report.pdf",
    page=1,
    dpi=150,  # 36–600
)
# out["image_base64"]  → "iVBORw0KGgo…"
# out["mime"]          → "image/png"

Set return_bytes=True to get image_bytes (raw PNG bytes) instead of base64 — useful when piping directly to disk.

AnthropicOfficeMcpBridge — Anthropic official skills via MCP

Configure the MCP server command once (any of the three methods):

# 1. Env var
export ANTHROPIC_OFFICE_MCP_CMD='npx -y @anthropic-ai/office-mcp'

# 2. Concinno credential store (persisted)
python -c "from concinno.core.credentials import CredentialStore; \
    CredentialStore().set('anthropic_office_mcp_cmd', \
    'npx -y @anthropic-ai/office-mcp')"

# 3. Pass server_cmd= kwarg directly per call (no persistence)

Then invoke:

from concinno_skills_office_advanced import AnthropicOfficeMcpBridge

# List what the server advertises
AnthropicOfficeMcpBridge().call(action="list_skills")
# → {"ok": True, "skills": [{"name": "docx", "description": "…"}, …]}

# Invoke a skill
AnthropicOfficeMcpBridge().call(
    action="invoke",
    skill_name="docx",     # docx / xlsx / pptx / pdf
    args={"title": "Report", "paragraphs": [...]},
)
# → {"ok": True, "skill": "docx", "content": ...}

If neither env / credential / kwarg is set, the tool returns {"error": "Anthropic Office MCP server not configured — …"} rather than silently swallowing the call.

Safety

• path must point to an existing local file under Path.home() or Path.cwd(). URLs (http://, https://, file://, data:) are rejected before any filesystem access.
• System directories (/etc, /Windows, /System, /Program Files, …) are blocked even when they happen to live on the same volume as $HOME.
• PdfAiExtract / DoclingPageImage also validate the suffix (.pdf).
• Files larger than 100 MB trigger a logger.warning — the tool still runs, but you get a stderr signal that cold-start latency will spike.
• AnthropicOfficeMcpBridge fails closed on unknown skill names (default allow-list: docx, xlsx, pptx, pdf; override via ANTHROPIC_OFFICE_MCP_SKILLS=docx,xlsx,…).

License

Apache-2.0. docling is MIT.