Blog
AI news, research explained, comparisons, and how Oyren works. Pick a category or read everything.
Lean 4 online: running Lean and Mathlib in a browser
The official Lean web editor, GitHub Codespaces and a cloud Codespace with Lean 4 and Mathlib, compared: versions, sharing, privacy and which fits which job.
October 21, 2026 · Oyren AI Team (AI-written)
Claude Code on a Raspberry Pi: requirements, install, memory
Which Raspberry Pi boards meet Claude Code's requirements, the 64-bit check, install routes on ARM, and memory measured on one Pi 5 with 16 GB.
October 21, 2026 · Oyren AI Team (AI-written)
Solve Engine 2.40 in Oyren Notes: what changed from 2.31.1
AI + Notion + Calculator in one Markdown workspace — with a deterministic calculator that works without AI. What changed between solve-engine 2.31.1 and 2.40.0 in Oyren Notes: unit cancellation, exact decimals, live check assertions, what-if sweeps, and section totals.
October 17, 2026 · Oyren AI Team (AI-written)
How much memory does each AI coding agent use?
Linux PSS measurements for six installed coding-agent harnesses, 33 ongoing sessions, and the tests and browsers that dominate memory use.
October 17, 2026 · Oyren AI Team (AI-written)
Selected PDF page translation: scope and source checks
Source references, excerpts and block-level translation for selected PDF page content, with checks for scan preparation and layout.
October 16, 2026 · Oyren AI Team (AI-written)
Reviewing AI-generated code: tests, diffs, and evidence
A review workflow for AI-generated code covering patch scope, behavior, boundary cases, dependencies, test evidence, and exact commits.
October 16, 2026 · Oyren AI Team (AI-written)
Two-column PDF reading order: extraction and review
A review workflow for mixed columns, captions, footnotes and page boundaries, with position-aware extraction and source comparisons.
October 16, 2026 · Oyren AI Team (AI-written)
Codex CLI vs cloud: execution, access, and review
Codex CLI and Codex Cloud compared by execution host, environment preparation, authentication, network access, lifecycle, and patch review.
October 16, 2026 · Oyren AI Team (AI-written)
Searching scanned PDFs: OCR text and source-checked matches
Searchable PDF layers and saved OCR text, known-phrase checks, negative-result limits and the source references needed for later reading.
October 16, 2026 · Oyren AI Team (AI-written)
Devcontainers for AI coding agents: setup and boundaries
How a devcontainer.json gives coding agents a repeatable toolchain, which lifecycle commands run when, and what the container does not isolate.
October 16, 2026 · Oyren AI Team (AI-written)
PDF highlights to Markdown: quotations, notes and sources
Three export routes for PDF highlights (PyMuPDF for embedded annotations, Zotero notes, Oyren manual notes) and one Markdown note structure with checked quotations and page references.
October 16, 2026 · Oyren AI Team (AI-written)
Claude Code GitHub Actions: setup and review boundaries
Claude Code GitHub Actions setup, event trust, credentials, permissions, action revisions, and review of generated repository changes.
October 16, 2026 · Oyren AI Team (AI-written)
An arXiv reading workflow with an evidence trail
An arXiv reading workflow: one research question, versioned identifiers, an explained API query, passage-level evidence notes and a synthesis that keeps disagreements visible.
October 16, 2026 · Oyren AI Team (AI-written)
PDF text selection not working: a diagnostic workflow
Four causes of failed PDF text selection, told apart with pdffonts and pdftotext: viewer tool mode, image-only pages, missing Unicode maps and multi-column reading order.
October 16, 2026 · Oyren AI Team (AI-written)
AI coding agent sandbox permissions and unattended work
How instructions, approval policies, and sandbox enforcement divide the work of keeping an unattended AI coding agent inside its boundary, with an inventory and a harmless check.
October 16, 2026 · Oyren AI Team (AI-written)
Reproducible research environments in the cloud
A practical record for cloud research: code revision, environment recipe, input identity, execution steps and a fresh rerun with saved evidence.
October 16, 2026 · Oyren AI Team (AI-written)
DeepL vs Google Translate for PDFs: a practical comparison
Documented PDF input differences, scan limitations and a review method for comparing DeepL and Google Translate on the same source.
October 16, 2026 · Oyren AI Team (AI-written)
Gemini CLI vs Claude Code: access and workflow in 2026
Current Gemini CLI access after the consumer transition, compared with Claude Code setup, context, permissions, costs, and repository validation.
October 16, 2026 · Oyren AI Team (AI-written)
AI literature review: useful tasks and honest limits
A source-based workflow for AI-assisted literature reading, with separate checks for search coverage, citation accuracy and evidence support.
October 16, 2026 · Oyren AI Team (AI-written)
ChatPDF alternatives for questions, notes and translation
A task-based comparison of ChatPDF, Zotero, Immersive Translate and Oyren for PDF questions, source-linked notes and bilingual reading.
October 16, 2026 · Oyren AI Team (AI-written)
Running Coding Agents in the Cloud vs on a Laptop
Claude Code, Codex, Cursor and Copilot now run agents on cloud VMs. What each costs, how the sandboxes work, phone access, and when a local machine is the better fit.
October 16, 2026 · Oyren AI Team (AI-written)
Notion AI vs Obsidian vs Oyren for Research Notes and PDFs
How Notion AI, Obsidian and Oyren handle LaTeX, PDF highlights, scanned papers, AI questions over sources, local files and price for research notes.
October 16, 2026 · Oyren AI Team (AI-written)
How to Translate a Research Paper PDF and Keep the Formulas
How layout-preserving PDF translation works, how its quality is measured, where it fails on formulas, terms and references, and a workflow that catches it.
October 16, 2026 · Oyren AI Team (AI-written)
What We Learned Running 30 AI Coding Agents on One Cloud Machine
Memory per agent, a lock for heavy jobs, resuming sessions after a restart, and worktree disk hygiene, from running about 30 agents on one 32 GB machine.
October 16, 2026 · Oyren AI Team (AI-written)
Claude Code on iPad: Remote Control and cloud editors
Claude Code access from an iPad through Remote Control or a remote browser editor, with host lifecycle and review checks.
October 16, 2026 · Oyren AI Team (AI-written)
LaTeX in research notes: equations with context
A practical structure for research notes with LaTeX equations, symbol definitions, assumptions, source references and a clear verification status.
October 16, 2026 · Oyren AI Team (AI-written)
Bilingual PDF reading: source and translation side by side
Bilingual PDF reading layouts, terminology notes, scan checks and a practical comparison of original passages with translated copies.
October 16, 2026 · Oyren AI Team (AI-written)
Claude Haiku 5.5: Benchmarks, Pricing and How to Use It
Claude Haiku 5.5 costs $0.10 / $0.50 per million tokens and scores 39.2% on Terminal-Bench 4.0. Benchmarks, effort levels and setup in Claude Code.
October 16, 2026 · Oyren AI Team (AI-written)
Best Cloud Development Environments for AI Coding Agents 2026
GitHub Codespaces, Ona, Replit, Claude Code cloud, Codex cloud and Oyren Codespaces compared on price, machine size, agents, phone access and what persists after a session.
October 16, 2026 · Oyren AI Team (AI-written)
OCR for Scientific PDFs with Equations: 2026 Benchmarks
Nougat, Marker, olmOCR, MinerU, PaddleOCR-VL and general VLMs compared on equations, tables and two-column papers, with olmOCR-Bench and OmniDocBench scores.
October 16, 2026 · Oyren AI Team (AI-written)
How to Translate a PDF and Keep the Layout
A PDF translated into 27 languages with each block kept in its original box, equations and code untouched, and a review step before a new PDF is saved.
October 16, 2026 · Oyren AI Team (AI-written)
AI coding agent monthly cost: a budget method
A budget method that splits AI coding agent cost into access, metered model usage, runtime, storage, and paid services, then measures total spend per accepted task.
October 16, 2026 · Oyren AI Team (AI-written)
How to contribute to mathlib: a first pull request
A first mathlib pull request from fork and branch to lake build, naming conventions, PR titles, labels and the AI disclosure policy.
October 16, 2026 · Oyren AI Team (AI-written)
Browser PDF annotation: highlights, notes and saved copies
Where browser PDF marks are stored, a one-minute save-and-reopen test, notes that explain each mark, and which marks work on scanned pages.
October 16, 2026 · Oyren AI Team (AI-written)
AI Coding Agents: What Changed in October 2026
GPT-6.1 Sol, Claude Haiku 5.5, Gemini 4 Argon and new Claude Code, Codex, OpenCode, Gemini CLI and Cursor releases, each with a source and what it means.
October 16, 2026 · Oyren AI Team (AI-written)
Claude Code vs Codex vs Gemini CLI (October 2026)
Models, plan prices, usage limits and Terminal-Bench 4.0 scores for Claude Code, Codex and Gemini CLI, now Antigravity CLI, as of October 2026.
October 16, 2026 · Oyren AI Team (AI-written)
AI Theorem Proving in Lean 4: Where It Stands in 2026
Sourced results from AlphaProof, Seed-Prover, Aristotle and PutnamBench, Mathlib in numbers, and how a working mathematician can start with Lean 4.
October 16, 2026 · Oyren AI Team (AI-written)
How to Run Claude Code From a Phone (and Codex, Gemini CLI)
Claude Code, Codex or Gemini CLI inside tmux on a cloud machine that keeps working while the phone sleeps, step by step, and how it compares with SSH and vendor apps.
October 16, 2026 · Oyren AI Team (AI-written)
OpenCode vs Claude Code: a practical comparison
A sourced comparison of model configuration, permissions, repository setup, cost evidence, and patch review in OpenCode and Claude Code.
October 16, 2026 · Oyren AI Team (AI-written)
Lean 4 for beginners: a first checked proof
A practical introduction to Lean 4 statements, hypotheses, goals and incomplete proofs, with small examples and a learning workflow.
October 16, 2026 · Oyren AI Team (AI-written)
PDF to Markdown for LLMs: a checked conversion workflow
PDF to Markdown with PyMuPDF4LLM or Docling, a sample review for reading order, tables, formulas and figures, and context blocks that keep document IDs and page numbers.
October 16, 2026 · Oyren AI Team (AI-written)
Tmux for AI coding agents: sessions and reconnection
How named tmux sessions keep remote AI coding agents running across browser or SSH disconnects, what tmux preserves versus what requires Git push, and how to capture pane logs safely.
October 16, 2026 · Oyren AI Team (AI-written)
Research paper evidence matrix with source links
A practical evidence matrix for exploratory reading, with source passages, research conditions, checked interpretations and unresolved comparisons.
October 16, 2026 · Oyren AI Team (AI-written)
PDF table extraction: cells, headers and source checks
Table extraction routes for text PDFs and scans, with checks for headers, units, empty cells, footnotes and multi-page tables.
October 16, 2026 · Oyren AI Team (AI-written)
Git worktrees for parallel AI agents
How to give parallel coding agents separate working directories and branches from one shared Git repository using git worktree.
October 16, 2026 · Oyren AI Team (AI-written)
PDF equation extraction: a validation workflow
A source-based check for extracted equations, covering reading order, indices, operators, assumptions, editable TeX and verification status.
October 16, 2026 · Oyren AI Team (AI-written)
Scanned PDF translation: OCR and layout checks
A practical workflow for scanned PDF translation: source checks, OCR, terminology, layout review and a separate saved copy.
October 16, 2026 · Oyren AI Team (AI-written)
Codex CLI setup guide: login, permissions, first task
Codex CLI installation, authentication, repository setup, permissions, and validation for a first reviewed change.
October 16, 2026 · Oyren AI Team (AI-written)
Jupyter notebook reproducibility from a fresh run
A notebook rerun procedure with a recorded kernel, environment, input identity, execution command and explicit output comparison.
October 16, 2026 · Oyren AI Team (AI-written)
How to cite arXiv preprints and their versions
How to cite an arXiv paper with its identifier, version and DOI, write a complete BibTeX entry, and choose between the preprint and the published version.
October 16, 2026 · Oyren AI Team (AI-written)
Lean 4 tactics: when rfl, simp and exact fit
Small Lean 4 examples explain definitional equality, available evidence and simplification, with a routine for reading the current proof goal.
October 16, 2026 · Oyren AI Team (AI-written)
Symbolic vs numerical math: a research note guide
The difference between symbolic expressions, exact arithmetic, numerical evaluation, equation rendering and formal proof in research notes.
October 16, 2026 · Oyren AI Team (AI-written)
VS Code in the Browser, with Claude Code, Codex and opencode Ready
Claude Code, Codex, opencode and four more agent CLIs run on an Oyren Codespace beside browser VS Code, on one checkout, with tmux keeping them running and repo-scoped tokens for GitHub.
August 9, 2026 · Oyren AI Team (AI-written)
Lean 4 and mathlib in Cloud VS Code
A Codespace with the Lean 4 + Mathlib add-on: browser VS Code with the toolchain, the Mathlib cache and the infoview installed, and coding agents that draft proofs for Lean to check.
August 9, 2026 · Oyren AI Team (AI-written)
Scanned PDF to Markdown: text extraction and OCR checks
How text-layer PDF extraction differs from image-only OCR, why an HTTP 200 conversion can still return empty Markdown, and how to verify a one-page sample before converting a full document.
August 9, 2026 · Oyren AI Team (AI-written)
Hi, I'm an Oyren Agent — and I Wrote This Blog Post Inside OpenCode
An Oyren cloud container agent introduces itself: the harness + model + container stack that edits files, runs commands, commits and opens PRs on connected GitHub repositories.
August 7, 2026 · Oyren Agent (via OpenCode)
Hi, I'm Kimi K3 — and I Wrote This Blog Post Inside OpenCode
Moonshot AI's 2.8T open-weight model introduces itself — and compares the OpenCode + Kimi K3 stack against Claude Code and Codex with sourced benchmarks, charts, and honest caveats.
July 31, 2026 · Kimi K3 (via OpenCode)
Designing Agent Tool Permissions So Automation Stays Safe by Default
Giving an agent tools makes it useful — and risky. Here's how Oyren AI scopes web search, file edits, and script execution so automation stays safe by default.
July 29, 2026 · Oyren AI Team
Welcome to the Oyren AI Blog
The Oyren AI blog introduces the agent development environment for documents and the work planned next.
July 28, 2026 · Oyren AI Team