automateyournetwork/document-generation

Document Generation

MCP Server

• Server: document-mcp (NetClaw-authored, spec 082 / roadmap R18)
• Command: $DOCUMENT_MCP_CMD
• Transport: stdio
• Credentials: none. This server writes files and touches nothing else
• Infrastructure: no device access, no ticket writes — so there is no approval gate here

The one rule that matters most

A document must never fabricate to fill a blank.

Tool output is ephemeral — read once, in context, by the person who asked. A document is not. It gets emailed, attached to a ticket, filed for audit, and read months later by someone who was not there, and it carries the authority of its formatting.

A professional-looking change record with a plausible invented number is a far more effective way to launder a guess into an official record than any amount of terminal output, because nobody re-derives a figure that is already in a table in a .docx.

So: never infer, estimate, interpolate, or carry forward a stale value to complete a document. If you do not have a figure, send {"unavailable": "<why>"}. The server renders that as visible text. It will not render a blank, and it will not let you send one.

Tagged values — the only accepted shape

Every value in a figure, table, keyvalue or form field is one of exactly three shapes:

{"v": "fgt-01", "src": "fgt_system_status", "device": "fgt-01", "as_of": "2026-08-03T13:58:00Z"}
{"unavailable": "device did not answer within timeout"}
{"failed": "FortiGate plane unreachable: connection refused"}

| You send | The document shows | |---|---| | {"v": x, "src": ...} | x, with a visible source | | {"v": ""} | (empty) — the source was consulted and returned nothing | | {"unavailable": r} | NOT AVAILABLE — r | | {"failed": r} | RETRIEVAL FAILED — r | | {"v": x} with no src | refused | | a bare x | refused |

unavailable and failed are different facts and render differently: a dead query is not an empty result. Neither ever renders as N/A, 0, -, or an empty cell.

Tools (6)

| Tool | Does | |---|---| | docx_write | Word document from ordered blocks: heading, paragraph, figure, table, keyvalue, image, pagebreak | | xlsx_write | Workbook from sheets of tagged rows, plus failed_rows | | pptx_write | Deck from slides: bullets, figure, image | | pdf_inspect_form | List a PDF's real named fields — call this before filling | | pdf_fill_form | Fill named fields into a new file; reports unfilled and unmatched | | list_documents | Find something you generated earlier |

What the server adds that you cannot omit

• A Source column on every table and every spreadsheet row — appended by the writer.
• An As of column carrying the source's own collection time, distinct from the generation time.
• Generation time and NetClaw attribution in the .docx footer on every page, the .xlsx banner, and the .pptx title and Sources slides.
• A Sources section in every file, listing each tool with its as-of and whether it was ok, partial or failed.
• A GAIT record for every call, including refusals.

If any figure lacks a source, the whole call is refused. There is no partial path.

Three limits, stated up front

No footnotes in Word. python-docx has no footnote API, so attribution is inline — a Source column in tables, a small parenthetical after prose figures. That is more visible than a footnote, not less. Word comments and document properties are set additively but never count as provenance: they are collapsed by default, stripped on paste, and absent in print.

No Office templates. .docx, .xlsx and .pptx are built from scratch and a supplied template is refused, not ignored — otherwise you would get an unbranded document believing it was branded. A corporate template's empty field is the strongest fabrication pressure in the feature. PDF forms are the exception because their fields are explicitly named: there is nothing to guess.

A filled PDF carries no Sources section. It is the customer's form and adding a page would alter it. For that one format provenance lives in the tool response and the GAIT record. Say so when you hand the file over.

Put figures in the right block

Prose carries no attribution. A paragraph asserting a bare number gets a caveat naming the block index — use figure, table or keyvalue instead, all of which force a source. Dates, ticket numbers, IPs, versions and "Section 2" are not flagged.

Other behaviour worth knowing

• Admin state and operational state must be separate columns. A merged status column is refused. An interface can be administratively up with no carrier, and collapsing them tells a reader traffic is passing when nothing is.
• Sources that disagree are both shown, with their origins and a caveat. NetClaw does not pick a winner.
• Untrusted text cannot become a formula. A description beginning with = is written as literal text — openpyxl would otherwise turn it into a live formula in an auditor's spreadsheet.
• Files are never overwritten. Output is timestamped and exclusively created, so a regenerated report cannot replace one already attached to a ticket. Path is reported back.
• ok means complete. A document with any gap comes back as written_with_gaps, and you cannot override that.

Boundaries — which skill owns what

| Want to… | Use | |---|---| | Draw a diagram | drawio-diagram, markmap-viz, uml-diagram, threejs-network-viz. This embeds their output and never redraws it — pass the file they produced plus src naming them | | Read a document into the knowledge base | rag-mcp (feature 062). It reads these formats; this writes them. Same libraries, opposite direction | | Create or update a change record | servicenow-change-workflow owns the CR lifecycle. This renders a document from one and never writes a ticket | | Send the document somewhere | slack-report-delivery, webex-report-delivery. Writing a file is in scope here; sending it is not | | Compose a NetClaw-shaped report | network-report-documents — the four standard compositions |

Important rules

1. Never fill a gap with a guess. Send unavailable or failed with a real reason.
2. Never send a value without src. It will be refused, and it should be.
3. Call pdf_inspect_form before pdf_fill_form so you map data to fields that actually exist.
4. Report what the response says. If it came back written_with_gaps, tell the operator the document is incomplete and which parts. Handing over a gapped document as if it were finished undoes the whole point.
5. Give the operator the path. They need to find the file.