Updates

Vicena Compute now supports Rowan-powered computational chemistry workflows through the Science Computer. Scientists can ask for descriptors, conformers, pKa, docking preparation, molecular dynamics setup, and related modeling work while Vicena builds notebooks, manages workflow state, and returns inspectable files, plots, tables, and structures.

Computation credits are now surfaced in plans and billing through a shared account credit view. Plus and Ultra include eligible compute usage, Free excludes Rowan and external simulation workflows, and the UI now shows remaining credits, monthly allowance, provider usage, reserved/refunded credits, protected interaction reserve, and the manual top-up path for larger AI, notebook, Rowan-powered, and future external-service work.

Dataset search and download tools now help Vicena locate public scientific datasets, resolve source-specific download roots, and bring selected data into the research workflow with clearer provenance and error handling.

The public computational chemistry page is now live with Rowan-powered workflow positioning, real docking visuals, plan and credit language, and clearer examples for molecular modeling, docking, descriptors, and simulation-oriented work.

Compute and external-service jobs now have a durable right-side job list for active and historical Rowan work. Job rows can show provider, workflow kind, input summary, submitted/running/completed state, status refreshes, result availability, credit usage, and clearer empty/loading/error states.

Scientific file viewing is now more native. JSON, spreadsheet-like tables, molecular structures, and Rowan result scenes can render directly inside Vicena with better inspection, editing, selection, and ask-about-this handoff flows.

Rowan job checking now includes a dedicated safe runner template and workspace skill guidance so existing workflow IDs can be polled without submitting duplicate paid jobs.

Long conversations now include a desktop breadcrumb rail for jumping between user turns. The rail appears only when it is useful, stays hidden on mobile, highlights the current viewport, and previews nearby turns on hover.

External API access is safer and more general. Sandbox tools receive scoped Vicena gateway tokens instead of upstream provider keys, existing sandboxes can refresh gateway credentials, and provider usage is routed through server-side ownership, credit accounting, and per-workflow controls.

Rowan workflows are better prepared for production use: default workspace skills, compact reference files, workflow quick references, sandbox snapshot support, gateway allowlists, fail-closed configuration, and per-user workflow ownership checks are now included.

Normal agent mode is now closer to Deep Work without becoming as autonomous. It can load bounded workspace context from /AGENTS.md, index user-owned .agents/skills entries, include relevant SKILL.md guidance when the prompt matches, communicate progress more clearly, and stop at a bounded iteration limit with a continuation prompt instead of silently looping.

OpenRouter-backed Codex work is more reliable across Qwen, DeepSeek, and other non-Codex models. Tool routing, reasoning-stream handling, duplicate thought suppression, and Deep Work model isolation have been tightened.

Normal and Deep Work now share the same Rust Codex core loop, tool lifecycle, and frontend event contract. Deep Work remains the more autonomous profile, while standard chats get the same core steering and live activity behavior when backed by Codex core.

Agent responsiveness is clearer during long turns. Tool activity starts earlier in the UI, streamed tool completion state is flushed more consistently, final loaders are hidden when a turn is done, and normal-mode activity no longer waits silently behind backend work.

Vicena now tracks the upstream OpenAI Codex Rust core more explicitly, with a stable upstream pin, documented integration boundary, and a guard script for checking whether the pinned Codex crates are current.

Agent activity in chat is now shown as a chronological workflow instead of scattered tool boxes. Preparation, live tool execution, notes, completions, warnings, and failures stay together so users can see both what is happening now and what already happened.

Chat readability has been refined with clearer contrast between user messages, assistant responses, and workflow activity, plus cleaner status badges, message-shaped workflow edges, and less visual noise around tool-heavy work.

Workspace and notebook UX received more polish, including better image rendering in chat, clearer notebook markdown/image guidance, computational chemistry tool catalog updates, and more useful public pages for Science Computer and AI chemistry workflows.

Scientific result viewers are smoother for real data review: molecule scene overlays and picking are more robust, spreadsheet and JSON renderers are easier to edit and inspect, and 3Dmol recovery is more reliable after notebook output changes.

Research link reading handles URL lists and extracted snippets more reliably, with stricter source handling and tests for the list-scanning path.

Workspace file links now open the intended workspace files instead of sending users to broken browser routes or page-not-found tabs.

Long, file-heavy, and tool-heavy chats now reload more reliably. Chat history serialization preserves rich tool outputs, files, and activity payloads without reopening the conversation into an empty page.

Rowan and external compute billing is more robust against runaway usage: workflow submissions reserve authorized credits, charge actual usage when available, enforce per-job/provider limits, and fail closed when required gateway configuration is missing.

Completed Rowan workflows no longer stay stuck as refresh-unknown or submitted in the compute jobs UI after a status sync. The gateway now persists workflow status updates, keeps sync best-effort when one provider check fails, and uses provider/object IDs and ownership scope when updating job rows.

Rowan batch workflow status and credit accounting are persisted more reliably across gateway reads, sync calls, and result polling.

Queued steering and active-turn controls are no longer limited to Deep Work copy. Standard Codex-backed turns can be steered through the same interface, and stale loading indicators are less likely to remain after the stop button disappears.

Codex Core is now a first-class Vicena agent runtime. Users can connect ChatGPT or Codex access, start Codex-backed chats, stream reasoning and tool activity, and keep coding-oriented agent work inside the same scientific workspace.

Planning Mode now lets Codex propose a work plan before acting. Plans can be accepted from the composer, revised through feedback, or declined when the agent needs to change direction.

Codex can now ask structured follow-up questions in the same response area as the composer. Recommended choices, numbered options, keyboard navigation, custom answers, and proceed actions are handled without starting a separate chat flow.

The model workspace has expanded into one surface for GPT, Claude, Gemini, DeepSeek, Llama, Mistral, Perplexity, and Codex-style agents while preserving files, notebooks, tools, protocols, and chat history together.

Codex Mode can now work through the broader model workspace where supported, including OpenRouter-backed models, with pricing and credit measurement routed through the same model-cost accounting used elsewhere in Vicena.

Workspace and Science Computer file transfer is now bridged. Codex can move files between Workspace Files and the Jupyter-backed compute environment so generated data, notebooks, scripts, and reports are easier to continue across surfaces.

Agent behavior has been refreshed around explicit work plans, skill and workflow guidance, progress narration, tool-call activity, managed tool execution timeouts, and active-turn steering.

Scientist Mode is available as a dedicated chat mode with a more structured loop for long scientific work, evidence-first reasoning, and clearer progress through research, analysis, notebooks, and protocols.

Long-running background work is more dependable. Background jobs are less sensitive to tab closes, phone sleep, and transient websocket disconnects, and notification emails now use a retryable outbox instead of being dropped after a temporary send failure.

Jupyter and Science Computer sessions are steadier during real workflows. Notebook iframes stay mounted while hidden, sandbox ownership is scoped per user, validation failures keep the notebook visible, and notebook actions report clearer live work status.

Prompt queueing and steering are cleaner during active Codex turns. Follow-up prompts can queue behind the current run or steer an in-progress turn with clearer controls and ordering.

Background-mode runs no longer stop just because the browser connection briefly disappears when background execution is enabled.

Repeated sends no longer create duplicate new chats or overwhelm the interface when the same prompt is submitted more than once during a slow or unresponsive turn.

Codex plan approvals and question answers now submit through the Rust route correctly, avoiding the generic “could not submit that answer” failure.

Codex and model selection are more predictable: Codex-native models use the Codex backend, non-Codex models stay on their normal model route, and selected model choices are preserved across Codex-mode conversations.

Pending Codex action UI no longer shows an unusable text input when only a plan approval or multiple-choice answer is expected, and the proceed control now stays inside the same response panel.

Notebook links, active Jupyter panels, and tool-result messages are less likely to disappear, reload, or duplicate during multi-step scientific work.

The full authentication migration is now included in Vicena 3.3: account identity, activation, login, logout, password setup, password reset, migrated-account recovery, and account deletion now run through Vicena's own first-party auth system instead of the previous external auth layer.

Long scientific agent runs are easier to supervise. Tool preparation, notebook updates, shell work, reasoning notes, and intermediate progress stay inside the work process instead of spilling into separate chat messages.

Jupyter notebook work now gives clearer live feedback while code is being drafted, including notebook names, cell counts, code-cell counts, and line counts when available.

Final synthesis after tool-heavy work is more dependable. Vicena now gets a lightweight completion nudge after successful notebook, shell, research, or workspace-file work, while preserving the agent’s ability to continue when more work is needed.

When an agent reaches its work-iteration limit, the conversation now reports that explicitly so the user can continue the run instead of seeing a silent stop.

Notebook content updates and streamed tool-call narration now render as traceable work-log events rather than duplicate assistant messages.

Completed work no longer keeps stale running indicators, pulsing dots, or outdated tool states after the final answer has arrived.

Research links, PDF handling, and generated workspace artifacts are more consistent when Vicena gathers evidence or saves files for later use.

Agent activity is easier to follow during long scientific work. Reasoning notes, tool preparation, shell runs, web research, notebook edits, and workspace file operations are grouped into one readable work process instead of scattered chat fragments.

Jupyter notebook workflows now show clearer progress while Vicena is drafting notebook code, preparing cells, running the notebook, and returning the resulting notebook path, outputs, and plots.

The Science Computer sandbox image has been refreshed for workshop use with updated scientific Python packages and preinstalled tooling for notebooks, single-cell analysis, imaging workflows, LC-MS/proteomics checks, and core-facility demos.

Final answers after tool-heavy work are more reliable. When a notebook, shell command, research pass, or workspace file operation has completed, Vicena is nudged to summarize the result instead of leaving the user with only a terse acknowledgement.

Workspace file creation and edits now return clearer saved-file outputs and stable workspace links, so generated Markdown notes, checklists, and research artifacts are easier to open again later.

Tool status is more precise: the interface now separates transport state from work outcome, so a command can show that the tool call completed while the underlying install, import, or notebook cell failed or was blocked.

Model compatibility around reasoning and tool-call narration is cleaner across provider formats. Intermediate planning is kept traceable in the work process while final synthesized answers remain separate.

The chat waiting state is quieter and more premium, replacing the old loading dots with a subtle animated ray and reducing visual noise during short response gaps.

Completed tool runs no longer continue to look active because of stale running indicators, pulsing status dots, or older tool-call state arriving after the final answer.

Right-side panel launcher icons now match the clearer white treatment used elsewhere in the workspace, improving visual consistency between the left sidebar and right rail.

Text chat routing and compatibility handling were updated for faster iteration, better fallback behavior, and more reliable handling of tool-capable model responses.

Free access copy now matches the actual product gates: Free is for academic or university email accounts, with research chat, literature, web, chemistry, and chemical analysis access. Compute, Workspace Files, persistent uploads, Protocol Audit certificates, sandbox/notebooks, and deep patent search remain Plus or Ultra features.

Plans now describe Plus and Ultra limits more accurately, including monthly credits, context windows, BYOK, Workspace Files storage, Science Computer access, Protocol Audit, patent search, priority models, and priority compute.

Literature Review now makes freshness handling clearer: indexed-paper evidence stays separated from current-literature web supplements, with source URLs, compact excerpts, and explicit evidence boundaries so the agent does not blur papers, supplements, and assumptions together.

Web Research now presents practical lab and engineering searches through clearer routing. Direct resource lookups, evidence-backed comparisons, supplier/spec checks, SDS/document searches, and broad research synthesis use more explicit evidence contracts and provenance.

Chat sidebar usability is cleaner across desktop and mobile: collapsed navigation keeps New Chat visible, the expanded New Research button now shares the same white new-chat glyph, and the narrow-layout control uses the same button treatment for a more consistent interaction.

Locked tools and models now use lock-only indicators instead of visible Plus or tier chips, preventing badge overlap on narrow screens while keeping upgrade state clear.

On narrow and collapsed chat layouts, the New Chat control remains available instead of being hidden inside the closed side panel, and its icon no longer shrinks or clips during the sidebar transition.

Workspace Files: Vicena now has a normal persistent per-user file system for saved protocols, audit certificates, Markdown notes, uploaded PDFs/images/CSVs, generated plots, and reusable research documents. Files use ordinary paths such as /Protocol Audits/audit.md and are separate from the Scientific Simulations/Jupyter filesystem.

File-aware agent workflows: the agent can create folders, write files, read existing documents, patch text files, move files, and return clickable workspace paths. Protocol Audit certificates and other durable outputs are now saved as files instead of being only transient canvas artifacts.

Protocol Audit V1 certificate workflow: protocol checks now produce structured audit certificates with formalized protocol structure, proof obligations, findings, assumptions, gaps, provenance, and readable rendered views.

High-level chemistry fact packs: Chemical Intelligence and Reaction Intelligence now consolidate PubChem, RDKit, Thermo, ChemPy, compatibility rules, energetics, kinetics, and model boundaries into structured evidence packets the agent can reason over.

Chat prompt queue: when an agent run is already active, additional prompts can be queued instead of discarded or interrupting the current run. Queued prompts preserve the selected tool, model, context filters, uploads, and Workspace File attachments, and can be edited or removed before they are sent.

Right-side work panels now behave like tabs. Files, Computer/JupyterLab, and Tools can be opened, switched, resized, and closed from the right rail without losing the chat, and file tabs support dirty indicators and drag reordering.

Workspace State: a read-only tool reports the current known JupyterLab UI notebook, the previous visible notebook, and the notebook most recently targeted by the notebook tool. This gives the agent a clear state contract without mutating files, switching tabs, or guessing from notebook history.

Public pages were added for the main Version 3 surfaces: AI for Chemistry, Protocol Design, Scientific Simulations, Workspace Files, Scientific Tools, and Updates. These pages can be read before sign-in and are listed in the public page registry and sitemap.

Protocol Design now has a dedicated public page. It explains the Builder + Audit workflow, where a protocol draft is treated as a claim to check against chemistry, physics, conservation laws, hazards, assumptions, and unresolved gaps before lab work.

Scientific Simulations now has a dedicated public page for the persistent cloud JupyterLab workflow: scientific Python, modeling, calculations, simulations, plots, computational chemistry, and notebook state that can continue across chats.

AI for Chemistry now has a dedicated public page that groups Chemistry Calculator, Chemical Intelligence, Reaction Intelligence, Protocol Builder + Audit, and Quantum Chemistry as one chemistry workflow surface.

Chemistry Calculator is now a bounded quantitative calculator for stoichiometry, yields, solution preparation, pH, simple equilibria, vapor pressure, and ideal-gas pressure. It validates required inputs and returns assumptions instead of guessing missing chemistry.

Protocol Builder now drafts evidence-grounded operational protocols with source-backed details separated from assumptions, missing critical details, and AI-led sanity findings before any lab use.

Research tools now use stricter evidence boundaries: Literature Review, Patent Research, and Web Research preserve source pointers, explain freshness checks, and keep scientific, patent, and general web evidence separated.

Scientific Tools catalog has been reorganized around how scientists work: Workspace, Protocols, Research, Chemistry, and Computer. Workflow-level tools are prominent; smaller atomic checkers are grouped under expandable concept sections with methodology, inputs, outputs, limits, tips, and examples.

Public landing pages now use a consistent hero prompt/send control and clearer navigation between Chemistry, Protocol Design, Workspace Files, Scientific Simulations, and the Scientific Tools catalog.

Chat scrolling is calmer during long runs. The chat keeps up with new messages when you are already at the bottom, but if you scroll up it stops forcing the viewport and shows a compact scroll-to-latest button.

Jupyter now uses an explicit action contract for open, create, read, append, insert, replace, edit, delete, move, run, and restart operations. New notebooks can be created with exact initial cells, create_notebook refuses to overwrite existing notebooks, and destructive actions require explicit cell indices or replacement cells.

Notebook outputs are returned as compact Jupyter transcripts with action, notebook path, cell_count_before/cell_count_after, changed cell source, execution counts, text output, error output, and image markers. Displayed notebook images are attached so the agent can inspect pixels when vision is needed.

Notebook-created support files now stay with the notebook. Relative file paths used inside notebook cells resolve beside the active notebook, so generated images, cube files, CSVs, and text outputs are easier to find and continue later.

Multi-notebook targeting is safer: notebook calls can explicitly target notebook_path, notebook_target="ui_active", or notebook_target="tool_active". When the browser-visible Jupyter notebook and the last tool-active notebook diverge, the tool can require an explicit target instead of silently editing the wrong notebook.

JupyterLab active-tab tracking is more robust. The browser bridge now observes the active JupyterLab notebook panel/tab, forwards visible notebook changes to the backend, and falls back to URL detection when needed.

Scientific Simulations session folders are now deterministic by conversation id and title, reducing collisions between chats with similar names while keeping notebooks and support files grouped by conversation.

Compute Shell and notebook support commands now report completion more clearly, so package installs, shell checks, and notebook support steps no longer appear to silently stop after a successful tool call.

The application backend was moved toward the Rust God architecture, with authenticated file routes, typed Python-server client integration, and the science tools extracted into the Python MCP server.

Workspace Files now keep each account’s saved files isolated to that user’s own workspace, with clearer handling for folders, document paths, and file operations.

Public pages are now checked for indexability as part of the frontend checks. Newly added landing pages are expected to return real HTML at their route instead of falling through to the app 404 before deployment.

The Workspace Files public page was adjusted for small screens so the hero copy, prompt, and workflow demo fit the viewport without horizontal overflow.

Protocol Audit: the AI now reviews your synthesis protocols for safety and feasibility before you go to the bench. Checks conservation of mass, yield limits, reagent-vessel compatibility, atom balance, boiling points, and GHS hazards — every finding cites its source.

Fail-safe chemistry tools: when the audit has no rule for a compound, it says so honestly instead of implying safety. Tools return facts and severity levels (critical / caution), never pass/fail verdicts — the AI synthesizes the final judgment.

Chemistry entity resolver: compound names are resolved to PubChem CIDs + canonical SMILES + structural classifications. Catches aliases (1,1,2-trichloroethane vs TCE), handles qualifiers (dilute / concentrated / solution), and matches chemistry by structure (SMARTS) rather than brittle string matching.

Scientific Tools catalog: a browsable side panel showing every tool the agent can use, grouped by category with descriptions, tier info, and ready-to-use example prompts. Open it from the "Browse all tools" button or from inside the tool picker.

Full precision in yield calculations: theoretical yields now report to 4 decimal places instead of rounding. A claimed yield of 13.9 g no longer passes against a true maximum of 13.858 g.

The AI is now labelled as the chemical safety authority in its system prompt: tools provide the exact data it cannot compute itself (molecular weights, boiling points, GHS hazards, compatibility rules); the AI provides the reasoning and context.

Tool picker now has a Browse link that opens the full scientific tools catalog without leaving your input.

Science Computer with JupyterLab: a persistent cloud computer where you and the AI agent collaborate in real Jupyter notebooks. Write code, run simulations, and visualize results together.

Multi-notebook support: create and switch between multiple notebooks in the same session. Each chat gets its own persistent workspace.

Session folders: each chat gets its own persistent folder with notebooks, data files, and scripts organized together. Files survive across sessions.

Computer control panel: welcome screen with power button, live preflight checklist during startup, idle timer, and power controls.

All models can now use tools: Gemini 3 Pro, Claude, GPT, and other models in the model picker have full access to research tools, notebooks, and the sandbox.

Tier-aware upgrade prompts: the AI knows your plan and guides you to upgrade when you need the Science Computer or other Plus features.

Scientific rigor: the agent now distinguishes formula visualizations from numerical calculations, verifies parameter magnitudes, and refuses to fabricate explanations when the mechanism is unknown.

Sandbox auto-stop: idle computers shut down automatically after 8 minutes of inactivity to save resources.

Updated pricing page with accurate context windows, storage tiers, and JupyterLab branding.

Quantum chemistry and molecular simulation: run DFT, Hartree-Fock, and MP2 calculations with PySCF, geometry optimizations with geomeTRIC, and atomistic simulations with ASE directly in the sandbox

Cheminformatics and structural biology: RDKit for molecular analysis, OpenBabel for format conversion, BioPython for sequence analysis, MDAnalysis for molecular dynamics, and py3Dmol for 3D visualization

Code Sandbox: run shell commands and Python code in an isolated Linux container that persists across messages in a chat

Full scientific Python stack pre-installed: numpy, scipy, sympy, pandas, matplotlib, scikit-learn, PubChemPy, Mendeleev, chempy, cclib, and more

Generated images (matplotlib plots, diagrams) and files (CSV, PDF, notebooks) are automatically detected and displayed inline in the chat

Download and upload files directly to the sandbox from the chat interface

Protocol Master integration: the simulation engine now executes actual computations in the sandbox for stoichiometry verification, thermodynamic calculations, and plotting

Sandbox compute is billed per-second based on actual resource usage, with automatic idle shutdown after 2 minutes and cleanup after 24 hours

Per-user container limits with automatic eviction of the oldest sandbox when at capacity

Improved security hardening across sandbox file operations

Browser tab now shows your chat title, making it easy to identify tabs when working with multiple chats

Background Mode: long-running research tasks now continue even after you close the browser or lose connection

Email notifications when a background task completes, fails, or times out, with a summary of results

Notification bell in the input bar shows unseen background job results when you reconnect

Per-chat BGM toggle to control which chats run in the background

WebSocket reconnection now seamlessly resumes in-progress background tasks without restarting

Hardened chat ownership checks across all API endpoints to prevent cross-user data access

Resolved race conditions in the background job registry that could drop or duplicate tasks

Wiki: publish any chat as a public, searchable page that anyone can read and share

Fork: continue someone else's published research in your own chat with one click

Publisher profiles: browse all published chats by a researcher via their public profile page

Display name setting so you can choose how your name appears on published work

SEO-optimized server-rendered pages with Open Graph, Twitter Cards, and Schema.org structured data for link previews

Dynamic sitemap and markdown endpoint for search engine and LLM crawler discoverability

16 ML-powered chemistry tools via a dedicated lab server: reaction prediction, retrosynthesis, molecule converters (name/SMILES/CAS), molecular descriptors, functional group analysis, similarity search, safety summaries, solubility checks, and spectroscopy predictions

Reaction prediction and retrosynthesis powered by ReactionT5v2 transformer models running on-device

Chemical formulas now render correctly as upright chemistry notation (e.g. H2O, NaOH) instead of italic math variables

Tool execution now shows a live status indicator while running, so you know the system is working

Tools no longer run twice for the same query due to improved context identification

MCP proxy hardened with response size limits, connection timeouts, and graceful error handling

Interactive landing page with a full walkthrough of Vicena's core research tools

Quick-start cards on the chat home screen redesigned for faster access to tools

SEO and discoverability improvements including sitemap, Open Graph, Twitter Cards, and structured metadata across all pages

Increased tool usage and lab simulation limits across all subscription tiers

Anonymous sessions now persist across browser restarts so you can pick up where you left off

Stronger security with token expiration and request validation

Resolved an issue where the first message in a new chat would not display

Sign-up flow now correctly handles sign-up state correctly

Free-tier users are no longer incorrectly blocked from using the platform

Chat title editing and chat suggestions work reliably again

Link reading tool restored for web-based research queries

Chat history no longer duplicates on reconnect

Comprehensive mobile UI overhaul — tighter spacing, better readability, and smoother interactions on small screens

Sidebar chats no longer require double-tap on touch devices

Lab View (Canvas) now opens fullscreen on mobile for distraction-free editing

Changelog page and version indicator so you can track what has changed

Chemistry Lab — Chemical Analyzer for properties, hazards, and reactive group lookup via PubChem & RDKit

Spectroscopy Predictor — functional group analysis with IR, NMR, and MS predictions

Reaction Kinetics Engine — rate constants, half-life, and time-to-completion calculations

Reaction Energetics Engine — virtual calorimeter for enthalpy and adiabatic temperature rise (thermal safety)

Solubility & Miscibility Engine — LogP, solvent compatibility, and partition predictions

Chemistry Calculator — stoichiometry, balanced equations, limiting reagents, vapor pressure, and ideal gas law

Redesigned tools UI with categorized research modules

Request-a-Tool card lets you suggest new research modules

Billing cost tracking and credit deduction accuracy

UI polish across the chat interface and overall visual consistency

Vicena branding rollout — new logo, colors, and design language

Research tool performance and depth optimizations

Staging mode configuration and URL routing

Agentic multi-step reasoning — the AI autonomously chains tools, evaluates results, and iterates up to 10 steps to solve complex problems

Protocol Builder — searches scientific evidence and drafts operational procedures with explicit assumptions and gaps

Lab Simulator (Protocol Master) — interactive simulation engine for testing protocols with feedback, world state tracking, and compiled output

Lab View (Files) — full-screen file editor for protocols, tables, code, and markdown

Thinking display — see the AI reasoning process in real-time as a collapsible block

Tiered credit system with per-tool cost tracking (Free, Plus, Ultra plans)

Bring Your Own Key (BYOK) — use your own OpenRouter API key

Multi-model support — Gemini, Claude, GPT, and more via intelligent model routing

Patent Research — patent database search with inventor and year filtering

Web search and literature review tools rebuilt for deeper, more accurate research

Product Researcher tool for market and product analysis

OpenRouter model routing with fallback and compatibility layer

Legal page with Terms of Use and Privacy Policy

Transition to GPT-4o model for better performance

Company and patent index integration for deep-tech discovery

Knowledge base via Google Drive integration

Updated to latest AI models with automatic fallback on rate limits

BHLI (Bio-Hybrid Living Interfaces) specialized research tool

Deep-tech outreach and campaign automation tools

Wide Research — expanded multi-source literature scanning

UI/UX improvements including auto-scrolling and copy buttons

Delete chat functionality

Migration to SvelteKit framework for better performance and routing

Stripe payment integration with flexible subscription pricing

Settings page with account management

Landing page with product overview

Literature Review tool — automated multi-paper research synthesis

Web Research tool — AI-powered web reading and summarization

Paper search with advanced filters (author, date, relevance)

Redesigned chat UI with new visual style

Chat renaming and sorting by last activity

OpenAI function calling integration for reliable tool use

Paper Search — scientific literature search powered by a custom Rust index

Multi-threaded search for faster paper retrieval

User registration and organization management

Switched to SQLite-backed chat history for reliability

Usage throttling to manage API costs

Multiple tool call handling in a single response

Production deployment on mirrorthink.ai with HTTPS

User authentication via Scitodate platform

Infinite memory via chat context management

Chat history and context awareness

Streaming responses with real-time token rendering

Graceful handling of long chats with automatic context window management

Initial prototype — Svelte + Tailwind frontend with FastAPI backend

AI chat powered by GPT-4 with real-time WebSocket streaming

First tool: YAML segment creation for the Scitodate platform