Official Convilyn client SDK — file conversion, agentic workflows, and a community library for reusable AI workflows.
Convilyn helps people turn repeated AI-assisted work into workflows that can be run again, shared, and improved. It is part of the CoreNovus vision for practical AI workflows across cloud services, local machines, AI PCs, edge devices, and future IoT environments.
Public mirror of Convilyn's monorepo (the source of truth). Contributions are welcome and land in the shipped package — see CONTRIBUTING.md (fork → PR → upstreamed, authorship preserved).
Convilyn is the official Python client for the Convilyn platform, and it has two halves.
Offline conversion runs entirely on your machine. convilyn.local works
with no API key, no account, and no network:
- Documents to Markdown — PDF, Word, PowerPoint, Excel, CSV, XML, plain text, and (with LibreOffice or Calibre installed) OpenDocument, legacy Office and ebook formats. Headings, lists and tables survive the trip, and embedded images come with it.
- Images between formats — PNG, JPEG, WebP, AVIF, TIFF and the rest of what Pillow supports on your machine. Transparency is composited when the target cannot carry it.
- PDF page operations — merge, select pages, split, rotate, compress, and add or remove a password.
Install only what you use (convilyn[pdf], convilyn[docx], convilyn[images],
…); plain text and CSV need nothing at all.
pip install "convilyn[docx]"
convilyn local convert report.docx --to mdRun convilyn local doctor to see what your machine can do and how to extend
it — it never guesses, and it never fails silently.
Platform workflows run in the cloud, and need a key. Convilyn(...) reaches
the hosted AI workflows — the part where a conversion becomes something you can
save, share, re-run and improve.
The long-term goal is for more of the first half to be possible: practical AI
workflows on the computing resources people already have, including AI PCs and
edge devices. convilyn-edge is where that work runs today.
Convilyn is for:
- Developers building reusable AI workflow components
- People working with file conversion and automation
- Builders experimenting with local-first AI workflows
- Communities that want open, practical AI tooling
- Contributors who want to help make AI workflows easier to use and share
uv add convilyn # or: pip install convilynPin the version. Do not reach for --pre:
pip install "convilyn[all]==<the exact pre-release you want>"pip already allows a pre-release when the specifier names one explicitly, so the
pin is enough on its own. --pre is a global switch — it applies to the
whole dependency resolution, not just to convilyn — and pip's own hint
(install with `pip install --pre` ) does not say so. Following that hint
pulls pre-releases of the packages underneath us too: pydantic, which every
model in this SDK is built on, and lxml, which python-docx / python-pptx
parse with. A result from that environment cannot be attributed to convilyn
at all.
That is pip's behaviour rather than a defect here, but anyone installing a pre-release will meet it.
See docs/QUICKSTART.md and the full docs at:
uv tool install "convilyn[all,mcp]" # or: pip install --user "convilyn[all,mcp]"
convilyn agent installInstalls a skill and registers an MCP server, so an assistant working in your
project can convert .docx, .pptx or .pdf itself. Safe to re-run, supports
--dry-run, merges into an existing config rather than replacing it, and writes
no credentials anywhere.
Install it where your editor can find it — the MCP server is started by the
editor rather than by your shell, and a project virtualenv is not on the
editor's PATH.
Each host looks in its own place, so both are written: Claude Code gets a plugin
at ~/.claude/skills/convilyn/ that loads with no marketplace and no install
step; Codex gets ~/.agents/skills/convilyn/SKILL.md plus an
[mcp_servers.convilyn] table merged into ~/.codex/config.toml.
To hand the same thing to a team from a marketplace instead, this repository is one:
/plugin marketplace add CoreNovus/convilyn-python
/plugin install convilyn@convilyn
The skill deliberately tells the assistant not to reach for convilyn on files
that are already text — .md, .txt, .csv, source code — where reading them
directly is faster and costs the same nothing.
docs/MEASURED-2026-08-28.md — the full test results, with the corpora, the metrics and the known limitations.
Every figure is produced by doc-eval,
a separate deterministic evaluator with no model and no network in its scoring
path, over named and versioned corpora you can re-run yourself. The filename
carries the measurement date, so an older report cannot be mistaken for the
current one.
Convilyn is maintained by CoreNovus, an independent developer team building open source tools for reusable AI workflows.
We welcome contributions from people who want to help improve documentation, examples, tests, file conversion workflows, local-first AI workflows, and developer experience.
A useful workflow should not disappear after one conversation. It should become something others can learn from, adapt, and run in their own context.
Start here:
CoreNovus is building toward a world where AI workflows can move beyond one-time chat sessions.
Convilyn focuses on the client SDK and community-facing workflow building blocks. The broader CoreNovus ecosystem aims to make workflows easier to create, save, reuse, share, and eventually run across cloud, local, and real-world environments.
We use the DCO (git commit -s) — no CLA.
Start with CONTRIBUTING.md, the Code of Conduct, and good first issues.
Small documentation fixes, examples, tests, and issue reports are especially welcome.
Please report vulnerabilities privately — see SECURITY.md.
Do not open a public GitHub issue for security vulnerabilities.