A research notebook designed by scientists, for scientists. Fermata keeps your ideas and your work at the center, with thoughtfully integrated AI tools that help you refine your research during every step of the scientific method.
Good science shouldn’t be rushed. Take your time to do it right, with Fermata.
getfermata.appFermata does its thinking through an agent that runs on your Mac under your own account: Claude Code with a Claude account, or OpenAI's Codex with a ChatGPT account. Set one up first and everything else is a double-click.
M1 or later. Check under Apple menu › About This Mac › Chip.
For Claude Code, a Claude Pro or Max plan or an Anthropic Console account. For Codex, a ChatGPT plan that includes Codex. Fermata's runs count against this account.
The command-line tool claude or codex. Install one with the steps below if its --version doesn't work in Terminal.
Open Terminal (press ⌘ Space, type Terminal, press Return) and paste:
curl -fsSL https://claude.ai/install.sh | bashThen start it once so it can sign you in. Follow the browser prompt, then type /exit.
claudeFermata is tested against recent versions of Claude Code. An old one still works, but each run shows a warning.
claude updateclaude --versionThe version should print as 2.1.250 or newer.
Open Terminal (press ⌘ Space, type Terminal, press Return) and paste:
curl -fsSL https://chatgpt.com/codex/install.sh | shThen start it once and choose Sign in with ChatGPT. Follow the browser prompt, then type /exit.
codexFermata uses Claude Code until you switch. Click your initials, top right, to open Settings, and under Models set Agent to Codex.
Fermata is tested against recent versions of Codex. An old one still works, but each run shows a warning. Run the install command again to update.
curl -fsSL https://chatgpt.com/codex/install.sh | shcodex --versionThe version should print as 0.162.0 or newer.
Download Fermata from getfermata.app. You get a file ending in .dmg.
Open it from your Downloads folder.
Updates only work from the Applications folder. If you skip this, Fermata offers to move itself there.
It is signed and notarized by Apple, so it opens without a warning. Startup takes a few seconds.
Click your initials, top right, to open Settings, and paste your ORCID iD under Profile. Fermata learns your field from your publications through OpenAlex (no account needed), so it won't ask you basics.
A run is one conversation about one question. Fermata never starts real work on its own. It waits for your go-ahead at step 3, and you can pause it at any point after that.
Type a question into the box on the home screen. That creates a project and its first run.
For a research question it asks one thing at a time and builds a summary. A simple question just gets an answer.
When the summary is ready, the play button starts the work. It is in the header and on the card that says it has enough.
One direction per idea, side by side. Each one reads, runs experiments and records findings. A checker tries to break every finding.
Every run ends with one, even a paused one: the answer so far, what each direction found, what's open, and the next step.
Tap an option or type your own answer. "None" is a fine answer and won't be asked again. To change the summary later, use the pencil in the header.
"Does adding a learning-rate warmup change final accuracy for small transformers on our dataset, or only how fast they get there?"
From the first question to the conclusion, the panel on the right of every run shows what Fermata is taking for granted, what it still doesn't know, and where the work could go. Check it as the run goes. Correcting an assumption early is cheaper than correcting a conclusion.
What Fermata is taking as given so it can proceed, in plain words. Each change is marked new, revised, merged or dropped, so you can see what moved since your last answer. Click one to question it in the chat, or edit its words.
What isn't known yet and would change the answer. Click one to raise it in the conversation. A question that blocks a direction is pinned above the feed instead, so you can't miss it.
The directions Fermata proposes, each with its hypothesis and what it would take. Open one to talk it through and shape its plan before you press play. Once running, each has its own card under Directions.
Tabs down the right edge of the run. Catch up retells a long run as a short list of scenes. Files holds everything the run wrote or read.
Anywhere in a run's conversation, or on a direction's plan, select text with your mouse. A small menu appears with three choices. You can also point at a figure.
Under Z-biased noise the threshold rises from about 1% to roughly 4.5% at a bias of 100, provided the decoder knows the bias.
Marks the words with a number and holds them under the message box, with an optional comment of your own. They go with your next message, so you can point at exactly what you mean. Select several passages to gather them into one message.
Writes a short note beside the words, explaining them as this conversation means them. The words keep a dotted underline. Click it to open the note again.
Ask one question about the selection and get one answer in a note beside it. For a follow-up, use the message box.
Notes are private and temporary.Break it down and Quick chat notes don't change the run, and Fermata's agent never sees them. They last until you quit the app. Copy anything you want to keep.
Write at any time. Your message waits in a queue above the box and is read at the next turn. You can delete it until it's read.
Stop ends the current turn only. Your next message carries on. Pause in the header also holds every direction until you resume.
Open the Assumptions tab and edit the one that's wrong, or click it to challenge it in the chat. Fermata reconciles the list on its next turn.
Open a finding to see its claim, where it holds, its limits and its evidence. Share it with the project when you're happy with it.
Paste a URL for a literature service or a skill. It installs, checks and turns on in one step, and is available in every project.
When a new version is out, an arrow appears in the header. Click it. Fermata downloads, swaps itself and reopens where you were. Your runs are kept.
Fermata looks for claude on your PATH and in ~/.local/bin, and for codex on your PATH and inside the ChatGPT app. Check in Terminal that claude --version (or codex --version) prints a version and that the tool opens without asking you to sign in. Check too that Settings › Models › Agent names the one you set up. Then quit Fermata with ⌘ Q and open it again.
Your Claude Code or Codex is outside the range Fermata was tested with. The run still works. Run claude update, or the Codex install command again, to clear the warning.
Fermata can only update itself from the Applications folder. Accept its offer to move there, or drag it there yourself and open it again.
Nothing is lost. When you open Fermata again, it resumes each run that was working and restarts the directions that stopped with it.
It is building the summary. Answer "none" or "that's enough" to skip a question, or press play as soon as the card says it has what it needs.
Open Settings, note the version, and click Open logs. Send the version, what you did, and the most recent log file to the person who gave you Fermata.
Fermata's agent runs through your Claude Code or Codex. It uses that tool's settings and permissions, and it can run commands on your computer the way the tool does. Experiments that go on the record run in a separate sandbox.
Your questions and files go to Claude or OpenAI under your account, like any Claude Code or Codex session. Runs and findings are stored on your Mac. Usage data counts clicks and screens, never what you type, and you can turn it off in Settings › Usage data.