Lemmaria

An AI writing agent for academics. Lemmaria works on LaTeX and Markdown papers the way a coding agent works on code — it reads your manuscript, proposes edits you approve one at a time, compiles and checks what it changed, and reviews your draft the way a program committee would.

Bring your own model key — Anthropic-protocol or any OpenAI-compatible endpoint, including a local one.

Nothing is written without your approval

This is the rule the whole extension is built around, not a setting.

• Quick actions (Polish, Proofread, Expand By, Shrink By, Translate) land in an inline diff you Accept or Reject.
• The agent's own edits arrive as proposals, per file, with a checkpoint you can restore from in one step.
• Every accepted edit is listed in the Edits view and can be rolled back individually.
• run_script never runs without your explicit approval, and only the interpreters you list.
• The review panel never writes at all.

There is no general shell and no arbitrary network access. Every capability the agent has is a named, bounded tool, and folder access outside your workspace is granted per session by you.

What it does

Write with an agent that knows the document. An agentic Assistant panel with sessions you can resume, fork, and rewind; conversation compaction when a session grows long; @ references to pull files and selections into context; and a command menu in the composer.

Review your draft before a reviewer does. Run blind panel in the Review tab (or /review in the composer) sends a blind panel of independent reviewers over your manuscript and returns a ledger of concerns, each one anchored to a byte-exact quote in your text — so every point it raises, you can jump straight to. Reviewers cannot see each other's findings, which is what keeps a panel from converging on one opinion. Lemmaria: Review Selection reviews just the passage you select.

See what a reviewer will see. With a model that accepts images, a PDF reviewer reads the compiled pages — layout, figures, page limits — beside the text reviewers, or on its own with Review the PDF (/review-pdf). A verifier re-reads every concern against your manuscript and withdraws the ones your text already answers, showing the passage that does. A scorecard then sums up the panel per dimension, computed from the concerns it located — never a prediction of acceptance.

Check what you cite. Lemmaria: Check Citations verifies your keys against your .bib and flags what it cannot ground. Literature search reaches Crossref and arXiv through a paced, polite client — with your contact e-mail if you choose to supply one.

Compile, and read the log. Lemmaria: Compile Check runs your TeX engine (latexmk, tectonic, xelatex, lualatex, pdflatex, or auto-detected), parses the log into findings you can navigate, and offers a fix for each.

Start from a venue. Lemmaria: New from Venue scaffolds a paper from a bundled venue template, carrying that venue's own writing conventions with it.

Teach it your conventions. Writing Skills are Markdown files — bundled, per-venue, or your own under ~/.lemmaria/skills — that Lemmaria loads as instructions when they are relevant. A ## Glossary table in PAPER.md becomes prompt context and is enforced on Translate.

See what it costs. Cumulative spend for this run, today, and in total, in the status bar or its own panel.

Your model, your key

Configure one or more models in the Models tab — Anthropic-protocol or any OpenAI-compatible endpoint, including a local one. Keys are stored in VS Code's secret storage, never in settings. Per-task ceilings on output tokens, tool calls, turns, and subagent fan-out are yours to raise or lower.

English or 中文

Every string Lemmaria writes renders in one language, chosen by your VS Code display language — Chinese for a Chinese editor, English for everything else. Your document text passes through untouched in whatever language you wrote it, and word counting picks its unit by script. If you read an English UI but write a Chinese paper, lemmaria.language pins the model's output language separately from the interface.

Getting started

1. Open a folder containing a .tex or .md paper.
2. Open the Lemmaria Assistant panel (Lemmaria: Open Assistant) and add a model in the Models tab.
3. Select a paragraph and try Lemmaria: Polish — you will get a diff, not a write.
4. Run Lemmaria: Create PAPER.md to give Lemmaria the standing context for your paper.

Requirements

• VS Code 1.106 or later.
• An API key for an Anthropic-protocol or OpenAI-compatible model.
• For compile checks: a local TeX distribution, or tectonic.
• For PDF review: a model that accepts images, and a compiled PDF of the paper — or a TeX installation to build one.

Status

Lemmaria is under active development, and its features may still change between versions. Questions and bug reports go to the project's GitHub Issues.

Licence

Lemmaria is licensed under the PolyForm Strict License 1.0.0: you may use it for any noncommercial purpose — your own research and study, or work at a university, a public research organisation, a charity or a government institution. The full terms are in LICENSE.txt, and third-party attributions are in THIRD-PARTY-NOTICES.md.