Core concepts
The workspace is organized around a research session. Each surface either adds context, generates output, or helps continue the work.
Sessions
Conversation, attachments, artifacts, and analysis runs kept together.
Artifacts
Abstracts, decks, and dashboards, plus inline charts, molecular structures, cited figures, rendered formulas, citations, and spreadsheets. Analysis runs add the files they produce; save a file to reuse it later.
Library context
Attach files to one session, click a name to preview them in place, or save them for reuse when the source matters later.
Agent memory
Follow-ups can refer to prior answers, exact cited source titles and links, and experiment outputs — including after a long session is compressed.
Keep researching while artifacts build
Abstract, presentation, and dashboard generation runs in the background. You can ask another follow-up immediately; Cortexa keeps the question and the artifact job progresses independently.
See when Cortexa is working
The active response has a moving light around its border while Cortexa is thinking or using tools. The signal stops as soon as the answer begins, and remains static when reduced motion is enabled.
Rate answers to reach the team
Every completed answer carries thumbs up/down next to Copy answer. A thumbs-down lets you pick what went wrong (facts, citations, instructions, depth, formatting) and add a note — that report opens a thread the team answers in My Feedback, exactly like feedback sent from the feedback pill. Ratings help prioritize quality reviews; they never change your session.
Copy answers into Word or Docs
Use Copy answer to place both formatted HTML and clean plain text on the clipboard. Word, Google Docs, Outlook, and other rich editors preserve headings, lists, links, emphasis, and tables; plain-text editors receive readable text without Markdown markers. The session Export menu offers the same formatted copy for the full transcript, while Markdown remains available as a download. Hovering any rendered table also reveals Copy table (pastes with structure into Excel, Sheets, or Word) and an Excel (.xlsx) download; dashboard table panels carry the same actions built from their source data.
Turn the latest answer into a deliverable
The most recent completed Cortexa answer keeps its abstract, presentation, dashboard, and experiment actions visible — even when the session ends with an unanswered follow-up or comes from an older saved session. Earlier generated artifacts remain attached to their original answers.
Run a Hugging Face model on your own data
For supported tasks, Cortexa can run an arbitrary Hugging Face model — inference only — against text or readable Library files as a background GPU job. You pick the model from a Hugging Face search; the result returns to the session and is saved as a reusable Library artifact with the model and run recorded.
You approve a hard maximum before anything starts
The card shows an expected cost and duration, the assumptions behind them, and a maximum charge. The model does not start until you approve that maximum. While it runs, the card shows live spend, progress, and a Stop button. Stopping ends the billing clock; reaching the approved maximum stops automatically and asks again before any continuation.
Estimates improve without learning from your data
Completed runs teach the estimator how long the same model and hardware take for inputs of a similar size. It records dimensions, timings, estimated cost, and actual cost — never the contents of your files. New or rarely used models show a wider range and a more conservative maximum.
Review the output before you rely on it
A model run is inference only: it runs the Hugging Face model you chose on the input you gave it and returns the result. Review that output before relying on it for regulated, confidential, or consequential work — running a model does not by itself establish accuracy or compliance with any standard.
Live web research
Cortexa can search and read the current web in a turn, or crawl a related set of public pages into a private, reusable Library snapshot.
Web pages and research databases work together
Cortexa uses live web search for current pages, organizations, news, and links you provide. For papers, trials, regulations, and structured scientific records, it prefers the dedicated research source because that evidence is cleaner, more complete, and easier to cite.
Crawl a public website in the background
Ask Cortexa to walk a public site when one answer needs many related pages. A crawl stays on the requested site, obeys
robots.txt, runs independently of chat, and saves up to 250 usable pages with direct URL citations and content fingerprints. It follows links up to 5 steps from the page you name, so anything buried deeper than that is not read. When the pages you want sit further in, start the crawl closer to them. A crawl also has 2 hours to finish, after which it stops and reports that it did not complete. You can keep chatting while it runs; Cortexa automatically continues the original request after the snapshot is searchable.Large crawls wait for approval
A crawl waits for your confirmation once it would cost about 50pages’ worth of credits, or sooner if that would be a tenth of the credits you have left. The card shows the page count and credit estimate before anything is spent. A crawl that starts small and turns out to be large pauses and asks again rather than spending the rest of your page budget. Final charges use the usable pages actually saved, not the estimate. You can cancel while mapping or crawling. Cortexa does not crawl login-only sites, bypass CAPTCHAs or paywalls, follow other domains, or use evasion techniques.
When a site is bigger than the crawl
No crawl reads more than 250 pages. When a site holds more than that, the card says how many pages it found and how many will be read, so you can see the snapshot covers part of the site before you approve it. On a site too large to survey up front, Cortexa assumes the full 250 pages and asks before it starts rather than guessing low. Point the crawl at a specific section when you need thorough coverage of one area.
Experiments
Experiment cells are for executable analysis. Use them when you need data preparation, statistical checks, chart generation, or a training pipeline handoff. Later cells can use earlier code and output summaries as context, but generated files are not automatically mounted into a fresh Python run unless they are saved to the Library or passed in explicitly.
Normalize cohort baseline
Experiment 1 · Python
Results
›Loaded 412 rows · dropped 17 with missing age/biomarker
›Saved figure → baseline.png
Naming experiments
Rename cells to match the analysis intent, such as “Normalize RNA-seq counts” or “Prepare training split.” Titles persist once the session can associate the cell with its saved analysis run.
Spreadsheet revisions preserve untouched cells
When you ask Cortexa to revise an existing
.xlsxworkbook, it uses a typed cell-edit operation instead of regenerating the file. The returned workbook is a new version with a cell-level change summary; if serialization changes an unrequested cell, the edit is rejected rather than published.Analyses that run for hours
Most code finishes inside the answer. When the work genuinely needs longer — a large parameter sweep, a long model fit, a simulation — Cortexa starts it in the background and finishes its reply instead of making you wait. A card in the chat shows progress, time used against the limit, and a Stop button. When the analysis finishes, Cortexa reopens the conversation with the results, so you can close the tab in the meantime.
What a long analysis is allowed
Unless the work asks for more, a background analysis gets 1 hour and is stopped at that point; 4 hours is the most it can ever be given. You can have 3 running at once, and past that Cortexa asks you to wait for one to finish, or stop it, before starting another. An analysis that dies unexpectedly is restarted 1 time and then stopped with an explanation, because a second identical failure is a fault rather than bad luck. These analyses are started from a chat session only, because they finish by reopening that conversation. The developer API, MCP clients, and Slack answer inside the request they arrived on, so they use the shorter in-turn analysis instead.
You approve the cost before a long analysis starts
Anything expensive — or anything that would use a large share of the credits you have left — pauses and asks first. The card shows the most it can cost and the longest it can run before you decide, and nothing runs until you choose. Those limits are enforced while it runs: an analysis that reaches either one is stopped and you are told why. You are charged for the processor time actually used, which is usually less than the estimate.
Python analysis
Experiment cells and the analysis Cortexa runs on its own both execute Python in a sandbox built for that run and discarded afterwards. This is what the sandbox already has, how long a run may take, and how many runs your plan includes.
What is already installed
Every analysis runs on the same Python build (
python3.13) with a standard scientific stack preinstalled, so a run does not spend its time fetching one: numpy, pandas, scipy, matplotlib, scikit-learn, statsmodels, requests, seaborn, plotly, kaleido==0.2.1, pyarrow, openpyxl, xlrd, beautifulsoup4, lxml, pillow, certifi, duckdb.Heavier readers install when your file needs one
The libraries that open EEG and MEG recordings, DICOM and other medical imaging, genomics files, and mass-spectrometry spectra are large, so they are installed for the run that reads your file instead of being kept in every sandbox. That costs the run a little setup time and leaves every other run unaffected. You do not have to name the library; the format of the file you uploaded decides it.
How long one run may take
A run defaults to 1 minute and can be raised to 5 minutes when the work needs it. Analysis of an uploaded dataset starts from a longer default, because staging and reading a large file happens inside the same window. Work that cannot finish there is started as a background analysis instead, under the limits described in Experiments.
How much computation you can ask for
Three sizes, and Cortexa picks between them from what you asked for. Inside the answer, a run gets up to 5 minutes on the standard sandbox. In the background, a run gets up to 4 hours on as many as 8 processors, and you can have 3 of those going at once. Anything that would cost real money stops and shows you the most it can cost before it starts. Billing is for processor time actually used, which is usually less than the estimate, and the limits are enforced while the work runs rather than checked afterwards.
What computation actually costs
A short analysis is a rounding error: a recent measured run finished in ten seconds and cost about a hundredth of a credit. The numbers worth knowing are the ceilings, because they are what you are agreeing to when you approve a run. The largest background analysis Cortexa will ever run — 8 processors for 4 hours — costs 360 credits, and with 3 running at once that is the most computation can draw from your balance at any moment. Approving a run reserves its maximum, so two runs can never promise the same credits. Nothing is kept running between your analyses, so idle time costs nothing at all.
Analysis runs included each month
One run is one Python execution, whether you started it in an experiment cell or Cortexa ran it to answer you. Each plan includes a monthly number of them, and the count resets with the billing month.
- Free · 10 runs
- Student · 20 runs
- Plus · 80 runs
- Pro · 200 runs
- Enterprise · 1,000 runs
Reaching the limit pauses analysis until the next billing month and leaves the rest of Cortexa working. See pricing for what else each plan includes.
Each run starts on a clean disk
Files one run writes are not on disk in the next one. To carry data forward, save it to the Library or pass it in explicitly, as Experiments describes.
Split panes
Split panes are designed as session-isolated views. Opening the same session twice will show the same saved state after refresh, but live keystroke-level collaboration is intentionally out of scope until the product adds real-time sync.
Up to 12 panes at once
Drag a session from the sidebar onto the edge of a pane to split it, in any direction, as deep as you like up to 12 panes in one workspace. Each pane runs its own session independently — a run continues whether or not its pane is the one you are looking at, and whether or not the tab is open. Panes get narrow well before the limit does, so on a laptop a grid of four to six is usually more readable than filling the whole budget.
Answer transparency
Cortexa tells you how well-grounded each part of an answer is, so you can see at a glance what needs a second look.
Clickable claim citations
Sourced claims carry an inline pill or source link — in prose and inside evidence-table rows, where each row's source column is clickable too. Hover or focus to preview captured evidence; open the source directly from the preview.
Evidence-completeness flags
When a cited source was only readable as the paper's own abstract or metadata (not full text), the citation is flagged inline — e.g. “(abstract only)”. On by default; toggle it under Settings → Preferences.
Computed figures show their work
A figure Cortexa worked out itself — a growth rate, a per-unit dose, a difference between two studies — isn’t in any source, so it arrives with the inputs it came from (each keeping its own citation) and the method, instead of borrowing a source’s citation as though that source published the result.
Projections are labelled as projections
A forward-looking figure is marked as projected wherever it appears — including inside tables, slide rows, and exports — and states the assumptions and the period it covers. Cortexa won’t project past that period, and won’t produce a forecast with no observed data behind it.
Scoped claims
Claims stay scoped to what Cortexa actually read. If retrieved sources don't contain a value, it says so rather than filling the gap from memory.
Evidence-first operational answers
For consequential procedures, doses, thresholds, deadlines, and safety factors, Cortexa prioritizes authoritative sources and cites the load-bearing claims inline. It aims to say when a value could not be grounded in what it read; that disclosure is not a hard guarantee, so check the citations on anything consequential.
How high-stakes questions are handled
Operational questions run an additional policy check that steers the turn toward authoritative retrieval, and cited sources pass a citation audit before they render. The check records whether the load-bearing value was grounded in a retrieved passage, and Cortexa is instructed to name what is missing when it was not — but it can still answer, so treat an uncited operational value as unverified. The policy check is a platform safety step and is not added to the turn's displayed credit consumption. Cortexa uses the recent conversation to distinguish consequential real-world instructions from nonclinical research analysis, including short follow-ups whose study context was established earlier.
When a turn reaches its processing limit
Cortexa stops starting new tool or analysis work, then answers from the evidence and artifacts already gathered. It distinguishes completed findings from remaining gaps, so the result may be partial without appearing to be a failed or silently truncated run. This is an automatic per-turn processing limit, not a budget configured on your account. Cortexa also stops after an empty tool result and one unsuccessful repair attempt rather than repeating searches that are not producing usable information.
When a tool keeps failing
A tool that fails the same way twice — an upstream that times out, a rate limit, a service that cannot be reached — is treated as unavailable for the rest of that turn. Cortexa stops calling it and says in the answer what could not be retrieved, instead of spending the turn on further attempts. Code and data-analysis errors get one extra repair attempt first, because the error message is itself something to act on.
Get full-text grounding
If a key paper is flagged as abstract-only, upload the PDF to the session or Library — Cortexa will synthesize from the full text and the flag goes away on the next answer.
Textbooks and open books
For long-form reference, Cortexa can search NCBI Bookshelf (clinical / biomedical textbooks), OpenStax (foundational STEM open textbooks), and DOAB (open-access scholarly books). Bookshelf results are free-to-read links and short snippets — not a full-text scrape. See the tool catalog for
search_ncbi_bookshelf, search_openstax, and search_doab.