Open Science changelog: what's new each month
Every stable Open Science release summarised by month: the highlights, new features, improvements, fixes and anything you need to act on. 5 months covered; the current month updates daily.
November 2026 · so far
5 releases: 1.0.7 → 1.2.1
November brought major expansions to Open Science's agent capabilities and research workflow tools. The month introduced cloud compute integration, a built-in agent runtime, mobile device pairing, and significant improvements to literature management and session handling.
- Built-in Open-Science Agent is now the default agent for new installations instead of Claude.
- Resource Center renamed and restructured; external resource links in Settings are now more compact.
- Compute settings have moved into the connections and resource management area.
Highlights
- Run compute jobs on Kubernetes, Hugging Face Jobs, Modal, and Ray with approval controls and recovery records.
- Export sessions as portable, narrated research films with terminal playback and voice narration.
- Access and manage the literature library from the command line and MCP clients.
- Pair mobile devices via QR code to open your workspace on phone or tablet.
- New built-in Open-Science Agent runtime is now the default for new setups, replacing Claude as the primary agent.
New
- Pi joins as a fifth agent framework alongside Claude Code, OpenCode, Codex, and CodeBuddy.
- Add references in batches of up to 100 from DOI, arXiv, PubMed, or article links.
- AI-powered snapshots summarize papers into five structured sections: research question, study population, methods, main findings, and conclusions.
- OCR for scanned PDFs is now on-demand with nine language packs available for download.
- Cloud backend support for Kubernetes, Hugging Face Jobs, Modal, and Ray with per-job approval workflows.
- Mobile device pairing via QR code to access your workspace over local WiFi or through auto-reconnecting tunnels.
- Optional resource recording on Windows to track process identities, session activity, and system memory usage.
- Scientific prompt optimization directly from the composer with review, edit, and undo options.
Improved
- Reading agents can now annotate PDFs by listing, locating, and creating notes in the literature library.
- Settings reorganized into focused tabs with individual descriptions for each scientific connector.
- Unified error surfaces for Composer failures with clear recovery actions for each failure type.
- Batch overview generation now resumes after interruptions without regenerating completed results.
- PDF reading order and extraction improved with better positioned text evidence and boundary verification.
- eQTL Catalogue queries restored to work with indexed files and GRCh38 assembly validation.
- Completed sessions release agent processes to reduce memory accumulation across all five frameworks.
- Macros update downloads only changed blocks on macOS, with fallback to full download when needed.
Fixed
- Update dialogs now display blocking process names instead of full system paths.
- Desktop recovery failures now retain actual error evidence in diagnostic packages and logs.
- Windows updates identify which files are holding the app open to assist with installation.
- Fixed-layout PDF translation now keeps formatting more reliably during exports.
- macOS verifies the running app before blocking an install to prevent update conflicts.
- Cloud compute sign-in for Modal and Hugging Face now respects system proxy settings for better network compatibility.
October 2026
14 releases: 0.34.1 → 1.0.6
October brought Open Science to version 1.0, adding project organization, session management, and a major expansion of scientific connectors. The month focused on research workflows, PDF handling, and reliability across desktop, web, and CLI platforms.
Highlights
- Version 1.0 release adds projects, conversations, skills, specialists, and connectors for organizing and conducting research.
- Session replay lets you step through recorded sessions and discuss them with your agent.
- Fifteen new scientific connectors arrived, covering genomics, proteomics, literature, imaging, and disease data.
- PDF translation is now resumable with original/translation/compare modes and better structural fidelity.
- Standalone Node backend runs open-science without Electron; desktop app connects as a native client.
New
- CELLxGENE Discover connector for searching public single-cell datasets with filtering by organism, tissue, disease, and cell type.
- Alliance of Genome Resources connector for cross-species gene search with orthologs, disease models, and phenotype annotations.
- Cellosaurus connector for cell-line identity and quality information.
- Monarch connector for phenotype association evidence across model organisms.
- PDC connector for cancer proteomics study and biospecimen discovery.
- ENCORI connector with tools for miRNA targets and RNA-RNA interactions.
- MCP server support in standalone CLI exposing research tools to external MCP clients.
- Zotero library import for bringing personal libraries and collections into the literature inbox.
Improved
- Notebook execution is now reviewable with per-run risk analysis before code runs.
- PDF structure extraction keeps figures with captions, reconstructs complex tables, and preserves native reading order.
- MaveDB functional scores, GWAS summary statistics, InterProScan sequence submission, and IEDB receptor searches expanded connector capabilities.
- New-conversation screen unifies research starters with the composer for better onboarding.
- GEO matrix discovery and preflight tools added to omics connector.
- Literature library now shows citation networks from Semantic Scholar with multi-source supplementary attachments.
- Skill selection now considers images attached to your message.
- Long conversations collapse intermediate messages while keeping running and waiting work visible.
Fixed
- PDF translation now works on Windows with drive paths and handles formulas next to tables without rejection.
- Desktop CLI works from non-ASCII installation paths including Chinese characters.
- Notebook kernels and environment captures no longer flash a console window on Windows.
- Large session and notebook snapshots transfer reliably without stalling or failing partway.
- Sessions paused for approval no longer flicker back to Running state.
- Artifact publication no longer expires during long-running turns.
September 2026
17 releases: 0.25.0 → 0.34.0
September brought major research capabilities to Open Science, including artifact editing, HPC cluster support, a literature reference library, and background task execution. The month also saw significant infrastructure improvements, new connectors for bioinformatics and data discovery, and a redesigned settings center.
Highlights
- Edit markdown, plain text, scripts, and source code directly as artifacts with version history preserved.
- Run notebook jobs on HPC clusters via Slurm or direct SSH, with durable submission and recovery.
- Import and organize a literature reference library with PDF support, duplicate detection, and citation formatting.
- Execute notebook, REPL, and shell jobs in the background while the agent turn completes immediately.
- Export sessions as portable .science packages with conversation branches, artifacts, notebook records, and environment locks.
New
- NVIDIA Build provider with Nemotron 3.5 Lightning and other curated agent-capable models.
- Slurm execution mode for remote compute hosts with durable job submission and polling.
- Literature reference library with collections, PDF imports, identifier-aware ingestion, and citation formatting.
- Artifact reproduction verification by replaying captured versions in isolated environments.
- Global workspace search across projects, sessions, files, and literature with grouping and filters.
- Verified skill marketplace with signed catalogs and one-click installation.
- Private session bookmarks for conversation passages and PDF selections with optional notes.
- Multiple independent side-chat tabs per session, each with its own preview and naming.
Improved
- Editable artifact versions with raw source editing for text, Markdown, scripts, and code files.
- Agent delegation now controlled per-session through a dedicated composer switch.
- App-managed Python and R runtimes can be reinstalled from Settings with optional automatic creation.
- Tool activity displayed as compact summary cards instead of raw JSON in messages and approvals.
- Batch PDF import with selection, destination choice, duplicate policy, and per-file progress.
- Settings reorganized into four purpose-based sections with workspace-wide search.
- Skill update previews rendered as readable diffs with line numbers and change highlighting.
- Provider connections validated before saving to catch credential or configuration issues early.
Fixed
- Protected notebook kernels on Windows launch directly under network sandbox without indirect paths.
- File exports now atomic, appearing only when fully written to prevent partial files.
- Skill catalog inconsistencies detected and repaired automatically on startup.
- Windows conda-managed R runtime restored with reliable startup and interpreter resolution.
- Notebook turns replay with same inputs and standard-library imports preserved.
- Generated image previews restored for agent-created images instead of showing placeholders.
August 2026
28 releases: 0.9.1 → 0.24.0
August brought major expansions to agent capabilities, collaboration features, and user control. The month added session management tools, persistent memory, cross-project navigation, and support for reading PDFs alongside existing code and data workflows.
- camelCase Host JavaScript APIs: snakecase methods like host.artifactPath are now camelCase (host.artifacts, host.agents, stopChild, sendFrameMessage, messageReceipt, resolveMessage, submitOutput).
- host.artifacts no longer accepts sessionId; callers must use frameId for exact producer filtering.
- Immutable invocation names for connectors and specialists: custom connectors must declare a lowercase-hyphenated invocation name at creation; displayName is editable but not used for skill generation.
- host.agents.list() now returns only id, name, displayName, description, and enabled; full editable records must be fetched separately.
Highlights
- Persistent agent memory lets agents remember important context across sessions within a project.
- PDF reading support enables agents to reference up to three linked documents directly in conversations.
- Session references allow you to link another session as context, giving the agent read access to its conversation.
- CodeBuddy joins three existing agent frameworks as a fourth selectable option.
- Project quick switcher lets you move between active projects without leaving your current session.
New
- Annotations: select text, images, or points in transcripts and previews to send as context to the agent.
- Persistent side conversations run independently alongside main threads with the same project and agent.
- Conversational skill creator lets agents draft and install new skills from natural-language descriptions.
- Session references (#) in the composer link to other sessions with turn-scoped read access.
- Notebook variables inspector shows the running Python or R namespace without starting a kernel.
- Centralized credential management stores GitHub tokens, API keys, and OAuth sign-ins in one place.
- Device-wide shared credentials can be bound to custom connectors as environment variables, headers, or OAuth flows.
- Network sandbox for notebooks and compute restricts outbound connections to approved domains.
Improved
- Live session status on Home shows running, waiting-on-you, and unread counts per project.
- Workspace project menu now previews other active projects for fast switching.
- Sidebar sessions group by live activity and date for easier navigation.
- Command palette searches across sessions, files, settings, and navigation within a project.
- Code syntax highlighting added for Python, R, JavaScript, TypeScript, CSS, and shell files.
- Specialist marketplace now includes signed package verification and CDN fallback.
- Context window dialog shows per-run composition with stacked bars and per-call usage details.
- Message center now displays project, task, and event preview context on notification rows.
Fixed
- Windows renderer crash recovery now triggers automatic reloads before falling back to manual recovery.
- Activity groups no longer require supplementary model calls, reducing per-turn API overhead.
- Live response rendering fixed to reduce flicker during streaming and survive view remounts.
- Session startup now reads query metadata from a materialized summary view instead of full file load.
- Composer message queue preserves message structure and order across view remounts.
- Permission denials now hold across the current turn, preventing agent retry or workarounds.
July 2026
20 releases: 0.1.1 → 0.9.0
July brought Open Science from early preview to a mature desktop application with multi-backend agent support, comprehensive notebook capabilities, and enterprise-grade reliability. The month saw major additions to storage, compute, and collaboration features alongside deep platform integration.
- Local Claude login replaced by isolated Claude subscription provider using setup-token.
- Legacy one-shot Python executor replaced by persistent runtime registry per session.
- Permission scopes now owned by Open Science (Once / This conversation) rather than ACP-specific lifetime.
Highlights
- Choose your agent backend: Claude Code, OpenCode, or Codex, swappable in Settings.
- Persistent multi-kernel notebook runtime supporting Python, R, and REPL with remote SSH compute hosts.
- Personal specialist agents with custom profiles, avatars, and scoped permission management.
- Native artifact versioning with full provenance tracking, checksummed content, and execution history.
- Message branching to explore alternative research paths without losing original conversations.
New
- Windows support with guided first-run onboarding and automated environment setup.
- Agent Skills management: create, edit, import skills as zip files, and invoke with / selector.
- Life-science data connectors aligned to upstream MCP servers (200+ tools across BioMart, literature, cancer models, and more).
- Multi-provider model configuration with built-in vendors (Claude, DeepSeek, GLM, Kimi, MiniMax, StepFun, SenseNova, OpenRouter, Xiaomi MIMO, Volcengine Ark, Grok) plus custom gateways.
- Browser-based workbench UI with headless daemon backend, manageable via open-science CLI.
- Remote SSH compute hosts for routing notebook execution to remote machines.
- Headless task SDK and CLI (open-science Node.js package) for automation.
- Artifact export as standards-compliant .ipynb files and conversations as Markdown or PDF.
Improved
- In-place auto-update on all platforms via electron-updater, no manual re-download required.
- File uploads now support drag-and-drop, clipboard paste, and files up to 10 GB with streaming.
- Permission system redesigned with syntax-highlighted code preview and per-conversation grant scoping.
- Context-usage indicator now breaks down window by category (system, conversation, tools, connectors, skills).
- Notebook runtime now serializes cell runs on shared interpreter and prevents double-fire connector calls.
- macOS builds are signed and notarized; Windows reliability pass fixed PATH overflow, shell execution, and managed-Codex install.
- Model provider settings include per-model multimodal capability metadata and reasoning-effort control.
- Configurable data-storage location with guided migration and support for relocating large project data.
Fixed
- Session resume now reconnects and continues interrupted sessions instead of dropping them.
- Providers failing connection tests are flagged as unverified rather than silently deleted.
- API keys fall back to reduced-protection storage when OS keychain is unavailable.
- R package installs no longer replace the live notebook kernel.
- PDF pages now rasterize at on-screen resolution instead of rendering soft.
- Blank R figures, micromamba banner flooding, and PowerShell CLIXML noise eliminated.
Summaries are written automatically from the official release notes (full changelog ↗); check the original notes before relying on a detail. Open Science: pricing, features and alternatives · All changelogs