MolEX Changelog
New features, improvements and fixes in each MolEX release.
v0.2.0-rc6.16 ·
Conversations can now be organized into projects, with shared files, project notes, and citations from earlier discussions in the same project.
New
- Projects: Create, rename, and delete your own projects from the conversation sidebar; move conversations into a project, across projects, or back to ungrouped, start new ones inside a group, and deleting a project keeps its conversations, which return to ungrouped.
- Sidebar browsing: The sidebar no longer stops at the first 50 entries — page through all projects, search every conversation by project name or title, and pin favorites; collapse and pin preferences survive refreshes and browser switches.
- Project files: Browse and preview files from other conversations in the same project and save them as shared project files instead of uploading again; they support search, filtering, and versioned editing, and a running task keeps the inputs it started with even if the originals change.
- Project notes: Add a note to a project that every conversation in it reads each turn; edits take effect from the next turn, and the note can be created, previewed, edited, cleared, or deleted.
- Cross-conversation search: A new conversation in a project can retrieve and cite earlier discussions from the same project; clicking a citation previews the original text in place without interrupting your conversation.
- Usage stats: Your workbench now shows token usage, GPU and CPU model invocations, and workspace storage with 7-, 30-, and 90-day views; the task overview adds average durations and refreshes automatically.
- Resource quotas: Each account has a monthly GPU time allowance and a cap on concurrent GPU tasks; tasks beyond the limits queue with the reason shown, and your actual monthly usage is visible in settings.
- Skill commands: Type a slash in the input to open a skill menu; picked skills appear as badges you can delete in one stroke, a message can carry several, and the model runs with all of them.
- Feature cards: Tapping a card on the home page now attaches its demo file and fills in a starter prompt; the attachment is visible and removable, goes out with your first message, and swaps when you pick another card.
- More multi-omics formats: Multi-omics uploads now accept .rar, .npy, .npz, .maf, .gef, .gem, .xenium, .qptiff, .czi, .svs, and .ndpi files.
- Structure properties: The properties panel now switches between structure files in a conversation, with the viewer and contact map always on the same file; it lists eight metrics such as chain IDs, residue and atom counts, and molecular weight, marking missing data N/A.
Improved
- Omics approvals: Multi-omics analyses in a regular conversation no longer need approval and start right away; other model analyses in the same batch still do, and the batch begins once every approval is in.
- Spatial omics loading: With classic Visium data files, the model now rebuilds the expected directory layout and loads the data instead of failing on how the files were arranged.
Fixed
- Fixed sending or continuing a reply doing nothing after the login session expired; the login now refreshes automatically and the message resends once, with a clear expiry notice if it can't be recovered.
- Fixed a single transient model error making every later turn of a conversation fail until a new one was started.
- Fixed sends failing nonstop after the selected thinking level was retired; the default level now applies automatically with a prompt to resend.
- Fixed continuing an old conversation after switching models mid-way mishandling earlier reasoning and sometimes erroring; unsupported combinations now clearly suggest starting a new conversation, and resumes, replays, and tool retries keep the model the run started with.
- Fixed multi-omics archives with special characters in file names failing to read correctly, and identically named files in multi-sample archives overwriting each other.
- Fixed scheduled tasks losing their cross-conversation memory, or staying stuck running forever when a memory file was damaged; turns with no real output now retry and notify on failure, and paused tasks are no longer re-enabled automatically or run twice.
- Fixed the task panel stalling when the data connection was busy until the page was refreshed; it now retries and recovers on its own, shows short readable errors, and the busy notice clears itself.
- Fixed losing the project selection when switching work modes in a conversation created inside a project, and Design-mode conversations created from an uploaded structure not joining the selected project.
- Fixed known issues.
v0.2.0-rc6.15 ·
BoltzGen can now design from an uploaded template structure.
New
- BoltzGen template structures: BoltzGen designs can now start from an uploaded template structure, validated under the same rules as other structure files.
Improved
- Model switch notices: Switching a model or thinking level now shows a notice in the lower-right corner, consistent with other notices in chat.
Fixed
- Fixed follow-up questions failing with an error in GLM thinking mode after several turns or tool calls.
- Fixed BoltzGen-generated .cif structure files failing to upload and being unusable in Design mode; they now upload and display correctly with residue selection, and files with incomplete structure info are no longer reported as empty.
- Fixed uploaded CIF structures being rewritten to PDB and failing as template structures in chat; CIF-capable models now receive files with the original residue numbering.
- Fixed download buttons in replies missing for root-level files and filenames with spaces or Chinese characters; buttons now appear promptly for newly produced files, and external links no longer show a download button.
- Fixed files uploaded to multi-omics tasks appearing as meaningless numbers; they now show their original names, and task inputs can use workspace file paths directly.
- Fixed task durations counting queueing time and appearing inflated; they now measure only actual execution time, and tasks without recorded start and end times no longer show an estimate.
- Fixed multi-omics tasks occasionally staying stuck at "Copying" after data copy, or flipping back to "Copying" after completion.
v0.2.0-rc6.14 ·
Science data search is here: query 40+ sources like PubMed, UniProt, and AlphaFold, covering literature, structures, and omics, right in a conversation.
New
- Science data search: Search 40+ scientific sources such as UniProt, PubMed, and AlphaFold, covering literature, compounds, protein structures, pathways, and omics, right in a conversation; results include sources, summaries, and license info.
- Science file downloads: Download structure, sequence, and other files from sources such as PubChem, KEGG, UniProt, RCSB, AlphaFold, and Ensembl; downloads are checked for format and size.
- Model list: The chat and model settings pages now list the models enabled on the platform with their capabilities; thinking levels appear under each vendor's own naming and can be selected directly.
Improved
- Model picker: Model and reasoning strength are now a single settings menu in the input, and your choice is remembered; if you never pick one, the platform default applies, and retired selections fall back to it.
- File uploads: Uploads now accept more formats, including .h5ad, .bam, .vcf, .bed, .gtf, .rds, and .h5seurat, covering omics, archives, and Seurat files; unsupported formats are flagged immediately.
- Knowledge base: Indexes now show a status for whether they are usable or need a rebuild; after switching the embedding model, an index needs one full rebuild, and retrieval states clearly when the model is unavailable.
- Science file handling: The model now sees the full set of supported file formats, so it picks the right format more reliably when working with results from PubChem, KEGG, and other sources.
- Image and video analysis: When the vision model is missing or reconfigured, uploads fail with a clear message instead of silently switching to another model; queued tasks ask you to resubmit.
Fixed
- Fixed saving in Design mode failing repeatedly after the Copilot panel was collapsed, which also blocked switching design history or returning to chat.
- Fixed canceled omics analysis tasks still showing result files; failed tasks now keep the outputs of completed steps.
- Fixed known issues.
v0.2.0-rc6.11 ·
Conversations now run in parallel without interfering with each other, and tasks keep running when you switch away or refresh.
New
- Sequence display: Long amino acid or nucleic acid sequences in replies now render as a single-line code block with a copy button.
- Drag-and-drop attachments: Drag a file anywhere on the page to attach it in a chat; unsent attachments are kept separately per conversation.
- Context compression notice: When a long conversation triggers automatic compression, a progress notice appears in the chat and clears when done; the summaries stay behind the scenes.
Improved
- Parallel conversations: Run several conversations at once and switch or refresh freely; messages, approvals, files, and task states stay in their own conversation, and reopening returns you to the last one.
- Persistent runs: Tasks keep running when you switch conversations or refresh, and anything missed is filled in when you return; resubmitting a message no longer duplicates the run.
- Thinking indicator: While an answer streams in Agent mode, the red stop button gives way to a pulsing thinking indicator; you can still stop generation with the send button.
- Input height: The chat input now shows at most five lines and scrolls internally for longer text, resetting after sending or clearing.
- Simpler model selection: The Auto model menu was removed from the input toolbar; messages still use the default model automatically, just as before.
- Welcome page card: The multi-omics card no longer says "coming soon"; it now lists supported analyses, including spatial transcriptomics, single-cell transcriptomics, RNA dynamics and pseudotime, gene regulatory networks, cell-cell communication, and perturbation modeling.
Fixed
- Fixed failed sends lingering in the chat as if they had been sent; failed messages are now retracted, a new conversation returns to a clean empty state, and successfully sent history stays visible.
- Fixed the stop button halting only a single job; stopping now cancels all model computation started by the current task, without touching other conversations or earlier runs.
v0.2.0-rc6.10 ·
Higher per-conversation limits on model and tool calls keep long bioinformatics analyses from stopping midway.
Improved
- Call limits: Higher per-conversation limits on model and tool calls let long bioinformatics analyses finish in the same conversation instead of stopping early.
Fixed
- Fixed antibody and protein design tasks being rejected when an epitope constraint was combined with preserved flexible regions; both constraints now take effect in the same task.
v0.2.0-rc6.9 ·
Task progress and notifications now update in real time, with fixes for history failing to load after switching accounts and long replies stopping.
Improved
- Upload feedback: On the welcome page, the send button shows a spinner and is disabled while an attachment uploads, preventing double sends.
- Workspace browser: The file tree loads folders on demand as you expand them, so the workspace opens faster and smoother.
Fixed
- Fixed a security issue.
- Fixed switching accounts loading the previous account's conversation and breaking history; unsent drafts and pending attachments are now cleared when you switch.
- Fixed the conversation you were viewing being wiped when a send or reply failed; chat history now stays intact after an error.
- Fixed conversations breaking during long, dense replies; streamed answers are now more stable.
- Fixed task progress not updating in real time on the dashboard or in the current conversation, and stale task states appearing after quick switches; real-time connections now reconnect automatically, and pages no longer occasionally become unreachable.
- Fixed experimental structures (crystallography, NMR, SAXS) wrongly showing pLDDT confidence; they now show mean B-factor and resolution, while predicted structures such as AlphaFold keep their pLDDT.
- Fixed approval cards not responding to Confirm or Cancel and reappearing repeatedly; the structure generation tool now asks you to fold the sequence first instead of returning an error on plain-sequence input.
- Fixed resuming, replaying, and tool retries failing when an older conversation had missing or outdated model info; these actions now run on your current default model.
- Fixed known issues.
v0.2.0-rc6.8 ·
Fixed
- Fixed known issues.
v0.2.0-rc6.5 ·
The molecular design workbench is here: upload a structure, run FoldX mutations, and review ΔΔG for a complete single-mutation design loop.
New
- Molecular design workbench: View a protein structure with a design assistant alongside, and upload your own structure files. Pick a residue, run FoldX, review ΔΔG, and accept or reject candidates; design state is saved and restored.
- Structure tree workspace: In Design mode, load a structure and expand a tree on the left to show, hide, or delete derived components, representations, and candidates; the structure viewer supports one-click fullscreen.
- Answer feedback: Like or dislike a final answer, with your choice restored when you reopen a past conversation.
- Knowledge citations: Replies that draw on a knowledge base now show clickable citation markers with a source list so you can check the original text, without stray symbols during streaming.
- Image artifacts: Images generated in chat can be published as artifacts and viewed in the conversation, supporting PNG, JPEG, WebP, and GIF; artifacts appear as compact cards with paging and full-screen previews.
- Usage trends: A new Usage page in settings shows daily and weekly model usage trends with per-model filtering, and clearly marks unknown costs.
- Environment info: See whether you can use the sandbox environment and your resource limits, with common disk paths now mapped to your personal workspace.
- Knowledge base editing: Rename a knowledge base or edit its description instead of deleting and recreating it, with edit entries on cards, lists, and detail pages.
- Multiple tool services: Conversations can now use vetted tools from several external tool services at once, and tools from other services keep working if one has an outage.
- Structure prediction: RoseTTAFold2 regains multi-chain complex support with a longer retrieval limit, and ColabFold now folds every uploaded sequence instead of only the first.
- Live environment retention: Your personal environment is no longer reclaimed while idle during a session, and a grace period applies after you close all pages.
Improved
- Model selection: Pick from a centrally maintained model catalog with a clearer set of options; if your chosen model is unavailable on resume or retry, you are told clearly instead of a silent switch.
- Knowledge base answers: Answers draw on the full retrieved content instead of a fixed-length excerpt per source, and the sources listed match what was actually used.
- Environment status: Messages sent while your environment starts or stops now get a clear starting/stopping notice, and plain questions that don't need the environment still work during that time.
- Notifications and task status: Notifications and active task statuses update live and are recovered automatically after a disconnection.
- Plugin upgrades: Upgrades now do a full overwrite after confirmation with a clear warning that changes are not merged or backed up, and the previous version is restored automatically if an upgrade fails.
- Sign-in and sign-up: Empty forms now report each missing field and focus the first one; layouts no longer jump when errors appear, and the verification-code button sits beside the input.
- Global file search: Clicking a file result in global search now opens the workspace and locates and previews that file.
- Publish confirmation: Publishing a knowledge base now asks for confirmation first, so an accidental click no longer makes it visible to everyone.
- Agent strategy: When the agent keeps probing with the same tool inefficiently, it gets a nudge to switch strategy sooner, cutting wasted calls and waiting.
- Welcome page: A placeholder card for multi-omics analysis is now shown as coming soon, and the protein-ligand binding analysis card was removed.
Fixed
- Fixed a security issue.
- Fixed pressing Enter to confirm an IME candidate sending the message instead.
- Fixed a single oversized tool output filling the conversation context and making it unusable for a long time; overly long outputs are now truncated.
- Fixed a conflict error when refreshing or re-entering a conversation awaiting tool confirmation; it reconnects to the same run and keeps history. Also fixed history being cleared on a load error and failed clarification answers being mislabeled as canceled.
- Fixed the file editor silently discarding unsaved changes when closed, switched, or left; it now asks you to save, discard, or cancel.
- Fixed files created in a conversation being uneditable in the sandbox; canceled tool operations no longer keep running, and progress and error messages now show correctly in chat.
- Fixed several knowledge base issues: rapid clicks no longer create duplicate knowledge bases, overly long names now fail clearly with a 200-character limit, and chunk overlap settings larger than the chunk size are rejected with an error.
- Fixed several settings issues: restoring defaults takes effect immediately, leaving with unsaved changes asks first, and the profile page no longer fails to load for some accounts.
- Fixed known issues.
v0.1.0-rc5 ·
Knowledge bases and scheduled tasks are here: upload documents for retrieval and Q&A, and run jobs on a recurring schedule or once at a set time.
New
- Scheduled tasks: Run jobs on a recurring schedule or once at a set time; every run has its own history and conversation replay, and you get an in-app and Lark notification when it finishes.
- Knowledge bases: Upload documents for retrieval and Q&A, view and download full texts, and set each base public or private; others' public bases carry a shared badge and cannot be edited, with a redesigned list and detail page.
- Structure visualization: The sequence and 3D view are linked: click a residue in the sequence to highlight it in 3D and see its number; charts add distance and contact-map heatmaps with mode switching, adjustable thresholds, and image export.
- Settings: A new settings page gathers your preferences, saved section by section, with a prompt to keep or discard unsaved changes; your profile now shows your email and phone number.
- Model service keys: Save, delete, and verify keys from the settings page; a saved key is only ever shown in part.
Improved
- Model parameters: Every adjustable parameter now has an explicit default and valid range that actually apply at runtime; the model list gains short capability descriptions, and parameter docs are more accurate.
- Parameter confirmation cards: Advanced parameters are now grouped and collapsible, groups with required fields open automatically, and hidden parameters no longer appear, making models with many parameters easier to browse.
- Bioinformatics models: The model hub has moved into workflow building, where models are picked from a dropdown and retired ones are greyed out and unselectable; names now display as generic categories, with a toggle to restore the original names.
- Small-molecule input: Accepts InChI, SMILES, and more formats such as .sdf, .mol, and .mol2; structure files are recognized by content rather than extension, with fuller validation and clearer messages.
- Execution details: Each step of the plan and every tool call now shows the model's accompanying notes, rendered as rich text.
Fixed
- Fixed some conversations failing with errors on every turn after an abnormal model output and becoming unusable; affected conversations now recover automatically on next use.
- Fixed tool calls from some models appearing as plain text in replies, which prevented the tools from running.
- Fixed some CPU compute tasks failing to run because of startup errors.
- Fixed a security issue.
- Fixed a crash when uploading CIF structure files that contain only coordinates and no other records.
- Fixed PDB searches always failing when exclusion terms were used; searches can now run with only inclusion terms or an organism.
- Fixed invalid structure files or model parameters failing midway through a task; you now get a clear error before submitting.
- Fixed saved conversation memory being impossible to preview or download, unfinished thinking or tool cards briefly expanding wrongly after a new message, and the structure viewer showing a stale structure when none is loaded.
- Fixed known issues.
v0.1.0-rc4 ·
MolEX's first release: chat and agent conversation modes, with tasks resuming automatically after a disconnect.
New
- Conversation modes: Pick Q&A or agent mode per conversation; the choice stays fixed for the rest of that conversation.
- Agent replies: Agent responses render as cards with reasoning laid out on a timeline; expand any tool call to see its input and output, and failures are highlighted in red.
- Reconnect and resume: After a network drop or page refresh, running tasks resume on their own without repeating actions already taken; pending approvals stay grouped together and keep their status across reconnects.
Improved
- Sequence alignment: Alignment computes faster, so tasks that include it finish sooner.
Fixed
- Fixed Boltz2 structure prediction tasks failing to run.
- Fixed tasks failing when some ligand structure information was missing.