altr-oss 0.2.1

Open source doc / excel / ppt skills for LLMs - Groq, Ollama, vLLM, any OpenAI-compatible endpoint.

altr

Open source doc / excel / ppt skills for LLMs.

Claude has document skills. ChatGPT has them. Open-weight models don't. Point gpt-oss-120b on Groq (or any model behind an OpenAI-compatible API) at altr and it gains three tools it can call to produce real files:

Tool	Output	Good for
create_document	.docx	reports, guides, letters, meeting notes
create_spreadsheet	.xlsx	budgets, trackers, datasets - with formulas
create_presentation	.pptx	pitch decks, talks - with speaker notes

The model sends structured JSON through standard tool calling; altr validates it (Pydantic) and renders it (python-docx, openpyxl, python-pptx). Validation and renderer errors are fed back to the model so it can correct itself - including server-side tool-call rejections from strict endpoints like Groq, which are turned into feedback and retried instead of crashing the run. No code execution, no sandboxes - the model can only emit document content.

Output ships with a clean default theme - accent-colored headings, styled tables and header rows, and charts drawn in a colorblind-safe palette - so files look designed out of the box. Or bring your own brand template (below).

Install

pip install altr-oss

The PyPI distribution is altr-oss; the import and the CLI command are both plain altr.

Or straight from the repo:

pip install git+https://github.com/PasinduSuraweera/altr-oss

Quickstart (CLI)

export GROQ_API_KEY=gsk_...

altr make "Create a 6-slide pitch deck for a solar-powered drone startup"
altr make "Make a 12-month SaaS budget spreadsheet with formula totals"
altr make "Write a 2-page onboarding doc for new backend engineers"

Files land in ./output. Works with any OpenAI-compatible server:

# Groq (default)
altr make "..." --model openai/gpt-oss-120b

# Ollama, fully local
altr make "..." --base-url http://localhost:11434/v1 --model llama3.3 --api-key ollama

# vLLM / LM Studio / anything else that speaks chat completions
altr make "..." --base-url http://localhost:8000/v1 --model my-model

Generation runs at --temperature 0.3 by default - tool arguments are structured output, and sampling cooler makes small models dramatically more reliable at producing complete, schema-correct documents.

Render a JSON spec directly, no model involved (great for testing):

altr render presentation examples/pitch-deck.json
altr render spreadsheet examples/budget.json
altr render document examples/onboarding-doc.json

Use it as a library

from altr import OfficeAgent

agent = OfficeAgent(model="openai/gpt-oss-120b", out_dir="out")
result = agent.run("Create a quarterly report with a KPI table")
print(result.files)   # [PosixPath('out/quarterly-report.docx')]
print(result.reply)   # the model's final message

Use it as a skill in your own agent

Already have an agent loop? Take just the tools:

from altr import SYSTEM_PROMPT, get_tools, dispatch

response = client.chat.completions.create(
    model="openai/gpt-oss-120b",
    messages=[{"role": "system", "content": SYSTEM_PROMPT}, ...],
    tools=get_tools(),          # OpenAI-format tool definitions
)

for call in response.choices[0].message.tool_calls:
    result = dispatch(call.function.name, call.function.arguments, out_dir="out")
    # {"ok": True, "file": "out/report.docx"}  - or {"ok": False, "error": ...}

dispatch never raises on bad model output - validation and render errors come back as data you can hand to the model for self-correction.

What the model can express

• Documents: headings (9 levels), paragraphs with inline bold/italic/code, bullet & numbered lists, styled tables with banded rows, images with captions, whole markdown blocks, page breaks.
• Spreadsheets: multiple worksheets, styled header rows, fitted column widths, frozen headers, live Excel formulas (=SUM(B2:B10)), a detected Total row set in bold, and bar/line/pie charts built from the sheet's data - if the model doesn't say which columns to plot, altr infers the numeric ones (and leaves the total row out of the chart).
• Presentations: title slides, section dividers, bulleted slides with indent levels (plain strings work too), chart slides, full-width image slides, speaker notes. Common model slips are repaired: chart data on a mislabeled slide still becomes a chart slide.

Filenames from the model are sanitized to their base name, so output can never escape the output directory. Image paths must point at existing local files - the system prompt tells the model to only use files you mention.

Brand templates

Start every generated file from your own template so fonts, colors, and slide masters match your brand:

altr make "..." --docx-template brand.docx --pptx-template brand.pptx
altr render presentation deck.json --template brand.pptx

Custom .pptx templates must keep the stock layout order (0 title, 1 title+content, 2 section header, 5 title only).

Passing a template switches the built-in theme off entirely - your fonts, colors, and masters are used untouched.

PDF export

Pass --pdf to make or render to also export each created file as PDF. Requires LibreOffice (soffice) on your PATH. From Python:

from altr import to_pdf
to_pdf("output/report.docx")

Roadmap

• Recipe/preset library of reusable prompts
• Configurable theme (swap the accent color and chart palette)
• Chart axis titles and number formats
• Nested markdown lists and blockquotes
• Watermarks and headers/footers

Contributing

PRs welcome - see CONTRIBUTING.md. Good first issues: roadmap items above, or new block types for the document schema.

License

MIT © Pasindu Suraweera