Changelog

Agent & API

Changelog for the Neutropic agent runtime, skills and API.

v0.1.63

October 26, 2026

  • ImprovementYearly subscriptions receive the full monthly allowance: A yearly subscription is granted the same monthly tokens as a monthly one at purchase and at each renewal. Yearly subscribers who received the smaller allowance introduced earlier this month have the difference added to their balance.

v0.1.61

October 23, 2026

  • FixCanva connection no longer asks you to reconnect after a token expires: Browsing Canva designs, importing one or sending a figure to Canva could fail with "Canva rejected the access token — reconnect" when a long-lived connection's access token had expired. The REST path now refreshes the token once and retries, the same way the Canva agent tools already did; the reconnect hint remains only when a refresh is really impossible.

v0.1.60

October 22, 2026

  • FeatureCanva connector released: With Canva's app review approved, the Canva connector is enabled for every account — connecting, the design picker, exporting figures and decks to Canva, and the Canva agent tools that were previously limited to reviewers.

v0.1.59

October 21, 2026

  • FixGPT-6 Astra works again: Every turn on GPT-6 Astra had been failing on the first model call since the model was added, because Astra rejects tool calls on the API path the other GPT-6 models use. Astra now runs through OpenAI's Responses API — tools, images, JSON mode, reasoning on Deep effort and streaming included — and both Basic and Deep turns have been verified live.
  • FixLong literature turns on GPT-6 Luna no longer stop with a provider error: When a tool that had been retired for the rest of the turn was called again together with another tool, two results could be recorded under one call id and the next model call was rejected ("Duplicate value for 'tool_call_id'"). The runtime now keeps a single result per call.

v0.1.58

October 20, 2026

  • FeatureBilling account and invoices: A new billing endpoint returns the subscription summary, saved payment methods, Paddle customer-portal links and the payment history, and each completed payment can be rendered as an invoice PDF or HTML page in the company statement format (A4, four languages, supplier block, Paddle-as-merchant note). Store payments are listed read-only.
  • ImprovementA new card shows up right away: A payment method changed in Paddle's customer portal is reflected in the billing account immediately, and the zero-amount verification transaction it leaves is hidden from the invoice history.
  • Improvement"Usage tokens" wording: Invoices for token purchases are titled "Neutropic usage tokens", and the limit messages (weekly cap reached, tokens used up) say "buy usage tokens" instead of "buy a top-up".

v0.1.57

October 19, 2026

  • Fix"Keep subscription" on a plan pass could fail: Reserving a subscription to start when a pass ends failed for some accounts whose stored payment-customer id was not known to the payment provider. The reservation now proceeds without that id (checkout matches you by email), and the customer-portal and refund lookups degrade gracefully instead of failing.
  • FixStore receipts cannot cancel a plan granted another way: A plan granted by the Neutropic team or a plan pass is no longer cancelled when an old App Store or Google Play transaction is re-delivered; store purchases only ever cancel plans the same store granted.
  • ImprovementChat errors in plain language: When a turn fails, the chat shows a localized notice — try again, nothing is charged if nothing was produced, contact the Neutropic team if it keeps happening — instead of the exception text. Session-expired and provider-busy messages are unchanged.
  • ImprovementStore subscriptions can replace a plan pass: A Google Play or App Store subscription bought while a promotion or tax-invoice pass is active now takes the pass over from the purchase date (a pending web reservation is cancelled, nothing was charged for it) instead of being refused, so the store never charges you without changing your plan.

v0.1.56

October 18, 2026

    v0.1.55

    October 17, 2026

    • FixVery long chats lost their newest turns: Once a chat grew past about 1,000 stored messages, the chat view, share page, export and the model's own history only received the first 1,000, so the running answer and every later turn were missing and each retry started over from the old context. Messages are now read in pages, so long chats show every turn.
    • FixTurns no longer end early with a thin answer: When a tool had been retired for the rest of a turn (for example after the re-read limit) and the model called it again, every following step was spent re-processing that call without a model call until the step budget ran out. Such calls are now settled at once and the turn continues normally.
    • ImprovementEffect pooling in the literature toolset: meta_pool, which pools effects collected from the chat's papers, now belongs to the literature toolset so research tracks keep a single toolset.

    v0.1.54

    October 16, 2026

    • FeatureNext-step suggestions on every turn: When a turn finishes normally, the agent proposes two to five next actions from this chat's work so far, the researcher's profile and project memory, and a research-lifecycle order (literature → hypothesis → design → analysis → writing → review). Each suggestion either deepens the same work with a more specific request or advances to another skill; the API event carries the skill key so clients can run it directly. The research pipeline's own gap questions feed into the same list. Interrupted, refused or survey-waiting turns get no suggestions.
    • FeaturePersonalized home prompts: GET /home/prompts?locale= returns one ready-to-send example request per research skill, in the UI language, written for the researcher's field, stage and recent chats once enough is known; otherwise a concrete human-science example. Not billed to the user.
    • FeatureSurvey analysis and inter-rater agreement: A new survey-analysis skill groups the questionnaire, rating-scale and open-ended methods (SUS, NASA-TLX, UEQ, semantic differential, conjoint, text themes, sentiment, data quality), and a new inter_rater tool reports percent agreement, Cohen's κ (weighted for ordinal scales), Fleiss' κ, Krippendorff's α and ICC with confidence intervals and interpretation bands.
    • FixCFA and SEM report SRMR and standardized estimates: The cfa and sem tools now include SRMR and the standardized solution they had been describing but not emitting.
    • ImprovementGeneration standards follow your data: The agent reads the session's data files to decide which domain standards apply — a survey CSV no longer receives biosignal or HCI reporting rules — and new experiment-review and writing standards cover participant protection, instrument sources, attention checks, attrition and effect-size reporting.
    • ImprovementConsistent email copy and design: Welcome, subscription, cancellation, renewal-confirmation and marketing-consent notices and the nurturing series share one greeting, signature and sender style, and the consent notices use the branded mail layout.
    • ImprovementToken accounting at vendor cost: Every cloud model is now metered at its vendor's actual price, with no markup or policy pricing. One managed-token unit equals the same real cost on every model, so lighter models stretch the allowance further and the same turn costs less than before on every model.
    • FeatureTop-up purchases: One-time Paddle purchases of ≈ 6.5M, 19M or 100M managed tokens are credited to the promotion pool as soon as the payment completes — each purchase once, with the token amount determined by the price bought — and are revoked automatically if the payment is refunded or charged back. /billing/config lists the available bundles.
    • ImprovementWeekly cap: The per-day cap becomes a per-week cap (a quarter of the monthly allowance, Monday to Sunday in the account's time zone), computed from the account's actual allowance so yearly plans and passes follow the same rule. The limit response is now WEEKLY_LIMIT with the next Monday as its reset time; the Lite fallback uses the week's remaining tokens; promotion and top-up tokens are outside the cap.
    • ImprovementYearly allowance: Yearly subscriptions receive their own monthly allowance (Lite 10.5M, Pro 21M, Max 50M tokens) when they start or renew; monthly allowances and passes are unchanged, and existing yearly subscriptions keep their current allowance until the next renewal.
    • ImprovementTurn estimates: /billing/config returns the estimated tokens of a Lite, Basic and Deep turn for every model, so the composer can show the cost before a turn is sent.
    • FeatureUsage insights: Every turn now records what it produced — artifacts, verified citations, review outcome, tools and sources used — alongside its tokens. New endpoints return the usage for a period, project or chat grouped by task, project, chat, model and day (/usage/report), one row per turn as CSV (/usage/report.csv) and per-turn detail for a chat (/usage/turns); thumbs-up/down ratings are joined per turn.
    • FeatureStorage management: Endpoints list every upload and artifact the account can see, with size, chat and project, and delete the selected ones (example-project files are protected), so the Storage page can free space without opening each chat.
    • ImprovementRenewal confirmation (South Korea): When Paddle requires a subscriber's explicit confirmation before a discounted subscription renews at the regular price, the account is notified by email with the deadline, the plan page can open Paddle's customer portal, and reconciliation keeps the pending state in sync with Paddle.

    v0.1.53

    October 15, 2026

    • FeatureGemini 4 Argon: Google's new flagship model (released October 13) is available on paid plans and is now the Gemini model used for Deep-effort turns; thinking depth follows the effort level (Lite / Basic / Deep).
    • ImprovementAnnounced price changes apply on their date: When a vendor announces a dated price change for a model, token accounting switches on that date automatically.

    v0.1.52

    October 14, 2026

    • FeatureShare links for chats: A chat's owner can create a read-only link (no expiry, 7 or 30 days), list active links with their view counts, and revoke them. Anyone with the link reads the chat's messages, plan and artifacts through the same screen as the owner; nothing can be changed through a link, and expired or revoked links stop immediately.
    • ImprovementLink previews show the chat title: The title a shared link reports to messengers and browsers is the chat's own title, without internal tags.
    • ImprovementNo-expiry links on paid plans: Links without an expiry require a paid plan; 7- and 30-day links are open to everyone.

    v0.1.51

    October 12, 2026

    • FixReports and follow-up questions in your language: A research turn asked in Korean (or Japanese or Chinese) could save an English report and suggest English follow-up questions while the chat answer stayed in your language. It happened when the agent received an internal reminder mid-turn, for example after a failed search; the report, its figures and the follow-up questions now keep the language of your request.
    • FeatureSlides skill: /slides builds a presentation deck from the chat's and project's reports, figures, charts and sources — a title, agenda, sections, comparison and timeline slides, a references slide and speaker notes, sized to the talk length. The agent can edit individual slides afterwards and checks the deck for slides with too much text, numbers without a source and citations that do not match.
    • FeatureDeck downloads: A deck can be downloaded as a single HTML file with a built-in viewer, as PPTX (editable text, native tables and charts, diagrams as real shapes) or as PDF, and its first slide is used as the thumbnail.
    • ImprovementHouse style for decks: Decks follow the Neutropic look — 16:9, navy cover and closing, numbered section kickers, cards and callouts — with five accent themes; a logo you attach replaces the Neutropic mark. Free-plan decks carry a faint watermark and download as PDF only; paid plans get all formats without the watermark.
    • FeatureExport a deck to connected apps: Google Drive receives a Google Slides file with speaker notes, Dropbox and OneDrive the PPTX, Canva a new presentation, and Notion or Confluence an outline page.
    • ImprovementEmail audiences from the welcome survey: Email sequences can target people by their welcome-survey answers — role, field, stage, how they found Neutropic and the features they expect.

    v0.1.50

    October 9, 2026

    • FixNo more silent outages when a provider's credits run out: The turn continues on another model with a short notice, and the team is alerted immediately.
    • ImprovementBusy-provider retries: Rate-limited requests are retried longer before giving up.
    • FixVery long turns keep saving: A turn that runs longer than your sign-in token's lifetime no longer fails to save its report or figures near the end.
    • FixErrored turns are marked and not charged: If the model provider fails mid-turn, the chat shows where it stopped and why, the line stays after a refresh, and no tokens are charged for a turn that produced nothing.
    • FixTurns lost during a deploy: A question sent while a new version was starting is resumed when you return to the chat instead of disappearing.
    • FixReply language: Answers follow the language of your latest message rather than the first message of the chat.
    • ImprovementPlan steps: The agent can mark plan steps by number or wording, so fewer steps are wasted on bookkeeping errors.
    • ImprovementCompressed tables: Gzip-compressed CSV and TSV files are read as data tables.

    v0.1.49

    October 8, 2026

    • FixPortrait phone videos: A video shot upright on a phone no longer appears on its side in Insight reports — the cover image, face and skeleton thumbnails, scene strips and the upload preview now follow the video's rotation, as it plays in the browser.
    • FixFace tracking for one person: Faces are no longer split into several people when someone moves quickly or leaves the frame for a moment. Duplicate detections and faces that never move (posters, screens) are removed.
    • ImprovementSmall and distant faces: Face analysis finds small faces in full-body or distant shots by locating the head through body pose and reading it enlarged, and marks them as less certain.
    • ImprovementClaude Haiku 5.5 for quick tasks: When you pick Claude, quick steps now run on Haiku 5.5 instead of Haiku 4.5, at the same token use. It can read up to five times as much context in one go.

    v0.1.48

    October 5, 2026

    • FeaturePictures inside documents: The assistant sees every picture in an attached Word, PowerPoint, Hangul or OpenDocument file and the figure pages of a PDF, with no fixed limit — very many pictures are sent at a lower resolution.
    • FeatureMore formats: Legacy PowerPoint (.ppt), OpenDocument spreadsheets (.ods), MATLAB (.mat) and R data (.rds, .RData) are read. Excel results say when a sheet has merged cells or hidden rows and columns.
    • FixReplies in the language of your question: A question about a file with a Korean name is answered in the language you asked in, not in Korean.
    • FixFollow-up research questions continue the research: Clicking a suggested follow-up after a literature review continues the research instead of answering "there is no evidence graph", and chats outside a project can find their own sources and reports.
    • ImprovementTopic first in literature screening: The screening table ranks sources that match your research question first, by meaning rather than exact keywords, so highly cited papers on a neighbouring topic no longer top the list. Duplicates, cover pages and peer-review records are marked, and a new "topic fit" column shows how close each source is.
    • ImprovementDistribution-only Hangul documents: Copy-protected (배포용) HWP files are not unlocked; the assistant asks for the original or a PDF.
    • FixMessages in your language: Upload descriptions, push notifications and sign-in and account error messages are no longer always Korean.

    v0.1.47

    October 4, 2026

    • Fix`@`-attached documents are read: HWP/HWPX, Word and PDF files attached with @ reach the assistant with their text.
    • ImprovementAttachment reading across formats: Long documents are read to the end in parts, every Excel sheet is reachable (form-style sheets included), text in Korean, Japanese and Chinese encodings is decoded correctly, and legacy Word, PowerPoint, RTF, OpenDocument text, e-books, Stata, legacy Excel and Apple iWork files are read. When something is left out (pages beyond 400, scanned pages, pictures, formulas without saved results), the assistant is told.
    • FeatureScanned PDFs and images: PDFs without a text layer are read as page images, photos over the size limit are resized, HEIC/BMP/TIFF are converted, and images uploaded without @ are shown to the assistant and stay visible on follow-up messages.
    • ImprovementLocked files: Password-protected files get one clear message to remove the password and attach again; the assistant does not try to unlock them or look for the content elsewhere.
    • ImprovementSafer file handling: Very large text files and oversized document archives are refused before they are read.

    v0.1.46

    October 1, 2026

    • FeatureLaugh, brand colour, music mood and thumbnail cards: Four new Insight skills — insight-laugh, insight-color, insight-music and insight-thumbnail.

    v0.1.45

    September 30, 2026

    • FeatureHeart-rate, voice and posture report cards: insight-heart, insight-voice and insight-posture work like the facial emotion report — the numbers are measured by the tool and the assistant writes only the plain-language reading. Asking in your own words picks the right card.
    • FeaturePresentation, reviews, ad and A/B cards: insight-presentation (one rehearsal video → expression, voice and posture), insight-reviews (review sentiment from a file or text), insight-ad (reaction scene by scene) and insight-compare (two versions side by side).
    • FeatureShare links and guest cards: Insight cards can be shared through a public link with consent and an expiry, and a guest can make one card without an account.
    • ImprovementSays why a named model can't be used: When you ask for a specific analysis model the server doesn't offer, the answer explains why and what was used instead.
    • ImprovementFaster previews: Thumbnails and document excerpts are prepared as soon as a file is saved and cached, and videos, PDFs, audio, spreadsheets and more get previews.

    v0.1.44

    September 29, 2026

    • FeatureFacial emotion report skill: insight and insight-face make an easy-to-read emotion report card from a face video or photo. The numbers are computed from facial landmarks and the assistant writes only the plain-language summary from them. Asking in your own words (for example "make an emotion report from this video") also works.
    • FeatureGroup analysis: Up to six faces in a video or photo are each analysed, numbered from the left of the screen and compared.
    • FeatureReport as an image: Emotion reports are exported as a PNG image, and their list and example cards get a cover image.
    • ImprovementHEIC photos in face analysis: Face analysis reads iPhone HEIC photos and respects the photo's rotation.
    • ImprovementRicher facial expression results: Facial expression analysis now also returns emotion scores, the main expression over time (including neutral) and a positive/negative curve.
    • ImprovementClaude Sonnet 5.5 for standard tasks: When you pick Claude, everyday steps now run on Sonnet 5.5 instead of Sonnet 5, at the same token use.
    • FeatureFigma connector: Paste a Figma personal access token to let the assistant read your design files, export frames as images into a chat, and read or post comments with your approval.
    • FeatureSlack connector: Connect Slack to post results, share files and search messages in your channels, and bring a channel's recent messages into a chat, each step with your approval.
    • FeatureMiro connector: Connect Miro to let the assistant find and read your boards and add sticky notes, diagrams and comments to a board you name, with your approval.

    v0.1.43

    September 28, 2026

    • ImprovementIllustrations follow your request more closely: Before drawing, the chat model you picked turns your request into a detailed visual brief: which parts to show, what to highlight, colours for each structure, and exact counts. The image model then draws from that brief. Requests like "colour the four lobes" or "highlight both hippocampi" now come out as asked, and scientific details such as the 8 × 12 wells of a 96-well plate are drawn correctly. The brief is written on Basic and Deep.
    • ImprovementHigher-quality Deep illustrations: On Deep, illustrations are drawn with Nano Banana Pro. Its shapes are cleaner and its colours more even, and they stay close to the original when converted to vector.
    • ImprovementMore faithful vector conversion: Converting an illustration to editable vector keeps its original colours much better. Differently coloured regions, such as the lobes of a brain, no longer merge into grey, and fine texture is smoothed into clean shapes that are easy to edit.
    • FixPower and sample-size numbers are calculated: Questions like "is 30 people enough for a correlation?" are now answered with calculated power, required sample size and minimum detectable effect, including for correlations, instead of numbers from memory.
    • ImprovementAnswers sized to your request: A question now gets a direct, short answer instead of a report. Asking to change a sentence, number or heading edits only that spot in the document; the rest stays exactly as it was, and no extra searches, sections or review are added.
    • ImprovementFaster, more careful support answers: The support assistant now uses a faster model. It answers common questions in about two seconds and passes anything the documentation doesn't clearly cover to our team instead of guessing.
    • ImprovementFriendlier, more accurate support replies: The assistant replies more warmly and politely in every language, apologises when something went wrong, and finds the right help page more reliably when you describe things in your own words (for example "cancel" or "how much").
    • FixHubSpot shows the connected account: Settings now shows the email of the HubSpot account you connected instead of "your account". Existing connections pick it up after Refresh tools.
    • FeatureLinear connector: Connect your Linear workspace to read and search issues, projects and documents, and to create issues and comments from a chat with your approval.
    • FixNo duplicate approvals for connected apps: If the assistant adds a field a connected app's tool does not accept, the field is removed before you are asked, so you approve once and nothing extra is sent.
    • FixConnected-app requests stay light: Asking the assistant to do something in a connected app (for example "create a Linear issue for our pilot study") no longer triggers a full research run or an extra report. It does the action and replies with the link.

    v0.1.42

    September 27, 2026

    • FeatureSupport chat: You can message the Neutropic team from the website or the app and get replies in the same chat, with email notice when you're not looking. Conversations stay private to your account, and your research data and chats are never copied into them.
    • FeatureAI support assistant: Support questions can be answered instantly from the Neutropic documentation and our support notes, in the language you wrote in. Questions about your account, billing or errors, and anything the assistant isn't sure about, go to a person.
    • ImprovementDeleting your account removes support chats: Account deletion now also removes your support conversations and their attachments.
    • ImprovementMCP open to all accounts: The Neutropic MCP server is no longer limited to early-access accounts. Every paid plan can connect AI apps and create access keys right away. On the Free plan you can see and disconnect connections, and tools and access keys unlock when you upgrade.

    v0.1.41

    September 26, 2026

    • FeatureUse Neutropic from other AI apps (MCP server): Claude, ChatGPT, Cursor, Claude Code and other MCP clients can connect to Neutropic and use it as a set of tools. They can search literature, verify citations, browse your projects and artifacts, run statistical tests and power analyses, start or continue a skill as a background task, upload files and save notes. Paid plans only.
    • FeatureSecure sign-in for connected apps: Apps connect with a standard OAuth sign-in and an approval screen. You choose which permissions (Read, Run, Write) and which projects each app can use. Every call runs under your own account's permissions, and disconnecting takes effect on the next call.
    • FeaturePersonal access keys: For Cursor, Claude Code and scripts, create a key with chosen permissions, an optional single-project limit and an expiry of 30 days, 90 days, 1 year or none. The key is shown only once.
    • ImprovementChats started from an AI app are labeled: Tasks an AI app starts run as regular Neutropic chats and remember which app started them.

    v0.1.40

    September 25, 2026

    • FixPlans finish cleanly: When a task ends normally, the step it was working on is now marked done, so the plan card no longer looks unfinished.
    • ImprovementSteadier live progress: Long quiet steps keep the connection alive, so the app can tell a slow step from a dropped connection and reconnect when needed.

    v0.1.38

    September 23, 2026

    • FeaturePlan passes: A promotion code can grant a paid plan for a set number of months. The pass ends on its own and the account returns to Free; accounts with an active card subscription cannot redeem one.
    • FeatureSubscription reservations for plan passes: A card subscription can be reserved during a plan pass — the card is registered at no charge and billing starts on the day the pass ends.
    • ImprovementLong chats just keep going: Each chat's context size now depends on the model rather than your plan. When a chat gets long, earlier turns are summarized automatically in the same chat, so you never have to start over and answers keep track of what came before.
    • ImprovementFewer cut-off answers: Every plan gets the same generous step limits, and a turn that nears its limit wraps up with a complete answer and a note on what's left.
    • FixGPT models with many connected apps: Chats on GPT models (including the free default) no longer fail with an error when many apps are connected. The most relevant tools are chosen for each turn so it stays within the model's tool limit.
    • ImprovementFairer daily limit on cheaper models: The daily allowance now accounts for each model's price, so free-plan chats on GPT-6 Luna keep the effort level you chose instead of dropping to Lite early.
    • FeatureNew models: Claude Opus 5.5, GPT-6 Sol and GPT-6 Luna: Opus 5.5 now powers Claude's deepest tier, and GPT-6 Sol and Luna replace the GPT-5.6 models. Opus 5.5 and GPT-6 Sol are available on paid plans; GPT-6 Luna is available on every plan.
    • FixGPT models can use tools again: Chats on GPT models failed whenever a turn needed tools such as search or file creation. They now run full agent tasks.
    • FixMedia, EEG and document tools restored: Facial expression and emotion, pose, heart rate from video, speech-to-text, EEG, PDF text for peer review, HWP files, and some app connections failed on the service. They work again, and if a tool is ever temporarily unavailable you get a short note instead of setup instructions.
    • FeatureHubSpot, Dropbox and Google Docs, Sheets and Slides: Connect HubSpot or Dropbox, or use Docs, Sheets and Slides through Google Drive: read your files, create documents, sheets and slide decks, and share links from chat.

    v0.1.37

    September 22, 2026

    • FeatureNotion connector: Neutropic connects to Notion's official server — search and read the pages you share, import pages and databases as data, and export documents to Notion, all under the same per-tool permissions and approval cards as Google Drive.
    • FeatureWorkspace pickers browse as trees: The Notion picker walks Private, Shared, Recently viewed, teamspaces, pages and database rows; the Google Drive picker shows your visible files and folders in their real structure.
    • FeatureZotero connector: Twelve tools on the Zotero web library — search, list and get items, collections and groups, CSL bibliographies (APA, Chicago …), BibTeX/BibLaTeX/RIS/CSL-JSON export, import of an item's PDF or metadata, and, with a key that allows writing, adding items, tagging, creating collections and saving this session's cited sources. Connects through Zotero's approval page or a pasted API key.
    • FeatureGitHub connector: GitHub's official MCP server behind the same permissions and approval cards as the other connectors — read and search repository contents, list commits and branches, write files, push several at once and create branches; the essential tools are exposed by default and the rest behind Expose all Standard tools. Exports without a branch go to the repository's default branch.
    • ImprovementGoogle Drive picker as a folder tree: The files and folders Neutropic can see appear in their real Drive structure; folders open but only files import.
    • ImprovementExport targets: The artifact panel offers Export to … only for connectors that can receive a document (Google Drive, Notion, GitHub).
    • ImprovementSmarter routing for material you point at: Asking to summarize, extract from or compare a Notion page, an uploaded file or an @-referenced artifact is treated as document work instead of starting a literature search.
    • ImprovementCancellation at period end: Cancelling a subscription is scheduled for the end of the current billing period; plan and credits stay until then, and a cancellation can be withdrawn.
    • ImprovementReview naming: The review skills are Review (report review · peer review) and Peer review; the peer-review document is headed "Peer review completed".
    • ImprovementLocal models: Prompts sent to a device model are trimmed where a small model gains nothing, and the model catalog recommends the 4B model on 16 GB machines.
    • FixLocal model turns: Long local turns no longer stall on unreadable binary content or on rate limits while streaming, and a tool error now triggers one retry before the model answers.
    • ImprovementMore steps per turn: Basic turns can take up to 40 steps (was 30) and Deep turns up to 64 (was 48); hypothesis discovery gets 32 steps at every level.
    • FixSources found in the same turn are checked immediately: Hypothesis, prior-work and source-verification tools now see the papers found earlier in the same turn — a fresh chat no longer reports "0 sources in session" and citations as out of range right after a search.
    • ImprovementVerify several sources at once: verify_source checks up to 10 sources in a single step, so verification no longer uses one step per paper.

    v0.1.36

    September 21, 2026

    • FeatureLocal models on the user's device: A desktop chat can run every model call of a turn on a local engine (llama.cpp, Ollama, LM Studio or an OpenAI-compatible server) with skill-scoped tools and no cloud metering; hybrid mode sends only auxiliary roles to the cloud.
    • FeaturePDF export of manuscripts and reports: Printed as the viewer lays them out (A4, page numbers) with figures embedded.
    • FeatureProtocol flow as a diagram: Experiment design draws the protocol flow with the figure skill — branches, merges and phases.
    • ImprovementGoogle Drive scope: Only files and folders picked in the Google picker (and files Neutropic created) are visible to the agent.
    • ImprovementWelcome and subscription e-mails: In four languages, with the Neutropic mail layout.
    • ImprovementE-mail confirmation: Password sign-ups must confirm their address before the API accepts their session.
    • ImprovementMobile task alerts: The turn-finished push respects the Task alerts on mobile setting.
    • FixFigure layout: Content is centred under a wide title.

    v0.1.35

    September 20, 2026

    • FeaturePeer review skill: /peer-review <journal> reads the attached manuscript (PDF, DOCX or Markdown), looks up the journal's scope and submission rules and its recent comparable papers, runs deterministic checks (rules, references, recomputed statistics), has five reviewer personas score and comment, and an editor issue the decision, a revision guide and recommended journals — calibrated so sound manuscripts are not rejected for style.
    • FeatureReview umbrella: /review picks the track — a report review of a document produced in the project, or a peer review when a journal or submission is in play; /report-review pins the former.
    • ImprovementReview, experiment-design and statistics reports cite their sources: Sentence-level [n] citations to the reviewed document, session literature and computation steps.
    • ImprovementCharts: Titles, legends, tick labels and axis titles are measured before drawing, so nothing overlaps or leaves the canvas.
    • ImprovementResearch tracks: Turning the Web track off really excludes web sources.

    v0.1.34

    September 19, 2026

    • ImprovementAnalysis and statistics reports cite their steps: Sentence-level [n] citations to this turn's analysis steps and datasets.
    • ImprovementTool pages: The first turn runs the page's skill; follow-ups are routed by intent with the chat's work so far as context.
    • ImprovementIdentity: The agent answers as Neutropic regardless of the model provider behind it.
    • ImprovementPaper writing: Citation styles and journal submission profiles are removed; the structural check (sections, length, citation coverage) stays.

    v0.1.33

    September 18, 2026

    • ImprovementUploads up to 1 GB: The upload limit rises from 200 MB to 1 GB, streamed so large files do not need extra memory.
    • ImprovementCoupon window: A registered discount coupon is valid for 24 hours.
    • FixStorage limit errors: A file rejected by storage for size returns a clear "file too large" response.

    v0.1.32

    September 17, 2026

    • FeatureArchives: zip, jar, tar (gz/bz2/xz/zst), 7z, rar and more are detected by content and unpacked with extract_archive; the files inside are readable by path.
    • FeatureAny file type: Uploads accept any extension; files without a reader are kept as attachments the agent can list but does not decode.
    • FeatureThinking in the transcript: Deep-mode reasoning is stored so a reload shows it where it appeared.
    • ImprovementFull-text reading: fetch_page returns the passages around the terms you ask for and serves open-access full text from Europe PMC when available; distinctive terms in a question are searched as exact phrases first.
    • FixKorean file names: Uploads with Korean or other non-Latin names keep their names.
    • FixDOCX export: Inline math becomes readable Unicode text; author names with particles (van der …) are cited correctly.

    v0.1.31

    September 16, 2026

    • FeatureDeepSeek and Kimi providers: DeepSeek and Kimi / Moonshot models, with their reasoning shown as thinking.
    • FeatureDiscount coupons: Coupon codes are registered in the app and applied at checkout or to a plan change.
    • ImprovementResearch reports: Comparison-table cells that the first pass could not fill are re-researched row by row; facts are looked up in Wikipedia and World Bank open data before web search; structured requests keep their exact part count.
    • ImprovementMind map and citation graph: The mind map shows the report's themes and keywords; the citation graph draws every cited paper and nothing else.
    • ImprovementFigure palettes: Colour schemes (Warm, Cool, Pastel, Earth, Vivid, Navy) separate from the drawing style; citation badges show the number inline before the label.

    v0.1.30

    September 15, 2026

    • FeatureHypothesis discovery, design level: <topic>_hypotheses.md now carries a prior-work table, a research-space map, established / contested / untested findings and 4–8 hypotheses each with rationale, novelty check, expected result and design sketch; hypotheses join the project evidence graph.
    • FeatureReview: what was improved: After a repair pass the review report opens with resolved, still-open and newly raised issues and the verdict before and after.
    • ImprovementLiterature screening: Each source carries its abstract and an extracted TL;DR.
    • ImprovementManuscripts in Japanese and Chinese: Section headings and word counts are recognised for CJK manuscripts; Korean manuscripts use a length scale matching Korean word counts.
    • ImprovementWriting turns: A manuscript that ends short is expanded section by section; figures are placed once and reused across writing turns.
    • ImprovementPublication date windows: "Before / since / between …" in a research request is enforced as a database-side filter.

    v0.1.29

    September 14, 2026

    • ImprovementNo clarifying surveys: Every skill decides from the request and records its assumptions instead of asking first.
    • ImprovementPaper writing: Tables and figures are placed inline where they are discussed, a Conclusion section is required, and manuscripts are written at full length section by section.
    • ImprovementAnswer language: Answers stay in the language of the request, including after context compaction.
    • ImprovementChinese transcription: Speech transcribed as Simplified Chinese.
    • Fix/compact: Works again after the language pinning change.

    v0.1.28

    September 13, 2026

    • FeatureDocs assistant endpoint: A public /help/ask route answers documentation questions from excerpts the docs site sends, streamed in the reader's language, with per-IP limits and a kill switch.

    v0.1.27

    September 12, 2026

    • FeatureDesktop local tool bridge: The hosted API orchestrates a turn while file, data and code tools run on the user's device through a local helper; device artifacts are registered as metadata-only rows.
    • FeatureLocal helper: A small FastAPI helper (health, capabilities, allowed-folder file access, sandboxed code execution, storage stats) shipped with the desktop app, with a downloadable media pack for video, audio and speech.
    • FeatureMobile push: Device registration and push notifications when a turn finishes.
    • ImprovementClient header: Requests carry the client kind (web, desktop, iOS, Android) for policy and analytics.
    • FixArtifact extensions: Artifacts created without an extension get one from their type; updates resolve the normalized name.
    • FixReview fallback: A review turn that ended with only structured JSON now renders a readable review report.
    • FixAudio-only recordings: Mobile recordings without a video stream are treated as audio.

    v0.1.26

    September 11, 2026

    • FeatureLive answer streaming: Model text is sent as it is generated; retracted text is cancelled with a retract event.
    • FeatureAudio and video tools: Transcription, acoustic features, voice emotion, facial expression, pose and rPPG from uploaded media.
    • FeatureEEG suite: Info, preprocessing, band power, topomap, ERP, ERD/ERS and SSVEP tools.
    • FeatureStatistical test selection: Assumption-driven choice between parametric and non-parametric tests with effect sizes and post-hoc tests.
    • FeatureGoogle Drive connector (MCP): Built-in server, per-tool allow/ask/block policy, in-turn approvals, encrypted credentials, file picker routes.
    • FeatureStorage governance: Per-plan storage quotas, a daily garbage-collection job and orphan cleanup.
    • ImprovementSkill prompts in English: All 17 SKILL.md files and tool descriptions unified in English; answers still follow the user's language.
    • ImprovementSlash aliases: Any top-level skill key works as a slash command.
    • FixStructured outputs: Structured JSON is produced only for the skills actually engaged in the turn.

    v0.1.25

    September 10, 2026

    • FeatureManuscript schema and export: Claim–evidence map, structured references, citation styles (APA, Vancouver and more), DOCX / LaTeX / styled Markdown export.
    • ImprovementWriting turns own their artifacts: /write delivers the manuscript and companion figures; gap searches happen in the writing turn without extra reports.
    • ImprovementEffort levels for writing: Lite / Basic / Deep stages with measured credit and quality differences.
    • FixDuplicate artifact events: The artifact summary no longer lists the same artifact twice after an in-place revision.
    • FixDesign figure consistency: Power curve, expected results and protocol flow are checked against the design report before saving.

    v0.1.24

    September 09, 2026

    • ImprovementArtifact language: Artifacts follow the request language more reliably.

    v0.1.23

    September 08, 2026

    • ImprovementIllustration skill: Quality tuning.
    • ImprovementPlan rename: The first paid plan is now "Lite" (was "Light").

    v0.1.22

    September 07, 2026

    • ImprovementIllustration skill: Quality tuning.

    v0.1.21

    September 05, 2026

    • ImprovementContext compression: Smaller tool observations sent to the model with less re-reading.

    v0.1.20

    September 04, 2026

    • FeatureScientific figure pipeline: Diagrams and charts are generated as vector figures from a Figure DSL; research and design figures moved from mermaid/PNG to editable SVG.
    • FeatureFigure editor API: Semantic operations and natural-language revisions on figures, saved as in-place revisions.
    • ImprovementFigure-only turns: Skip the reviewer when a turn only produced figures.
    • FixStray reports: "Draw a figure" no longer produces an extra report.md.
    • FixTruncated titles: Session titles are no longer cut off by thinking tokens.

    v0.1.19

    September 03, 2026

    • FeatureEffort levels: Lite / Basic / Deep policies control search caps, sub-agents, report length, review repair and step limits.
    • FeatureDaily allowance: Per-plan daily token caps with monthly grants.
    • FeatureContext compression layer: Type-aware reduction of tool results, logs, files and search results before they reach the model.
    • FeatureModel catalog 2026-09: Updated vendor models, prices and routing tiers.
    • ImprovementArtifacts readable by tools: read_file and list_files see figure, report and table artifacts.
    • FixSchema check: Trigger functions are skipped correctly.

    v0.1.18

    September 01, 2026

    • FeatureStructured artifacts: Analysis reports, hypothesis sets, manuscripts and review reports are editable and re-renderable.
    • FeatureHuman-in-the-loop approvals: Risky code (deletions, overwrites, participant exclusion, external transfer) pauses for approval.
    • FeatureR runtime: code_exec can run R scripts in the same sandbox.
    • FeaturePatent claim extraction: Claims parsed into structure from Google Patents pages.
    • FeatureIn-app plan change: Change plans without a second checkout — upgrades apply immediately, downgrades at the next renewal.
    • ImprovementStructured tool errors: Failures carry a code, retryability and a recovery hint.
    • FixBilling hardening: Fixed the subscription-activation failure from the 2026-09-01 incident.

    v0.1.17

    August 31, 2026

    • FeaturePydantic AI runtime: The agent loop's model calls, tool dispatch and streaming run on pydantic-ai while the scientific harness stays in-house.
    • FeaturePersistent kernel: One Python process per session keeps variables and DataFrames between code_exec calls.
    • FeatureScientific toolset: 13 deterministic tools (statistics, psychometrics, biosignals) added.
    • FeatureSub-agents and context management: Delegation to child agents and tiered compaction of long histories.
    • FeatureDurable execution: Turns are checkpointed at step boundaries and resumed after interruptions.
    • FeatureProvenance graph: Source → claim → artifact lineage derived from session records.
    • Feature/compact: User-triggered conversation compaction.
    • FeatureStudy workflow, quick fact, policy skills: /study, /fact, and the analysis/statistics/hypothesis/writing/review policy skills.

    v0.1.15

    August 27, 2026

    • FixEmpty report sections: Section synthesis retries when a model call returns no text.

    v0.1.14

    August 26, 2026

    • ImprovementResearch artifacts: Better research output structure.

    v0.1.13

    August 25, 2026

    • ImprovementChat titles: Titles summarize the conversation.
    • ImprovementResearch Markdown: Research artifacts in Markdown with restored highlights.

    v0.1.12

    August 24, 2026

    • ImprovementCredit balance: Token credits and usage are reported to the chat, the Usage panel and the plan page.
    • ImprovementSettings › General: Profile settings are read and saved through the API; your profile name is applied to chats.

    v0.1.11

    August 23, 2026

    • FeaturePaper table and ranking: Literature searches produce a screening table ranked by citation score, journal score and similarity.
    • FeatureE-mail: Sign-up and subscription confirmation emails.
    • ImprovementBackground review: Reviews run in the background.
    • FixDesign reports: Reference annotations and sample-size consistency with the power curve.

    v0.1.10

    August 21, 2026

    • FeatureExperiment design report: A fixed-structure design report with one clarifying round; figures → reference designs → report.
    • ImprovementContext-aware routing: The skill router uses conversation context, not just the latest prompt.
    • ImprovementMarkdown research reports: Reports switch from HTML pages to Markdown.
    • ImprovementTerminology: "Emotion engineering" enforced as the English term.

    v0.1.9

    August 20, 2026

    • FeatureServer-side allowance: Metered turns are gated by the caller's wallet.
    • ImprovementDesktop CORS: Bundled desktop app origins allowed.

    v0.1.8

    August 19, 2026

    • FeatureBilling: Checkout connected; pricing and token policies revised.

    v0.1.7

    August 18, 2026

    • ImprovementResearch skills: Search and report quality.

    v0.1.6

    August 14, 2026

    • FeatureProject memory: Chats in a project share context; artifacts from sibling chats are readable.
    • FeatureGemini 3.x: Catalog update with forced tool-call mode for models that write tool calls as text.
    • ImprovementArtifacts created box: Per-turn description of what was produced, with one-line artifact descriptions.
    • FixSource cap: The current turn's sources are no longer pushed out by the 60-source report cap.
    • FixGemini function calls: Restored the thought signature Gemini 3.x requires.

    v0.1.5

    August 12, 2026

    • ImprovementResearch skills: Search and report quality.

    v0.1.4

    August 11, 2026

    • ImprovementResearch skills: Search and report quality.

    v0.1.2

    August 06, 2026

    • FeatureResearch tracks: Paper, patent and web tracks composed by the research umbrella; search_patents tool.
    • FeatureDetached turns: Turns keep running after the connection drops; credits are charged server-side.
    • ImprovementVenue filter: Chosen venues are a hard filter on paper searches.

    v0.1.1

    August 05, 2026

    • ImprovementResearch skill: First round of improvements.

    v0.1.0

    July 28, 2026

    • FeatureNeutropic 2.0: Initial draft of the agent and API.