Ask Ray

Ray is the text box in the middle of the dock. Type what you want in plain English and Ray does it in the workspace — opens files, runs tools, writes and runs notebook cells, builds figures, sets up schedules — or answers from these docs when the question is about Relay itself.

⌘J puts your cursor in the bar from anywhere. Esc stops a run that is still going.

Ray in its own window

Drag Ray’s bar out of the dock while it is closed and it becomes a window you can move by its title bar and resize from any edge or corner. Right-clicking the dock and choosing Open Ray in a window does the same, as does Ray: Open in a Window in the command palette (⌘K). A reply being written, your draft and its attachments all come along.

To put it back, drag the window onto the dock, press Return to the dock in its title bar, or close the window.

What Ray can see

Ray can see whichever window you have in front, so this file, these cells and the selected region mean what you are looking at. It also knows the active Project and the files in your Data tree, so the kidney slide from Tuesday usually resolves without a path.

It does not read every file you own on every question. If you want Ray to look at something specific, attach it.

Ask Ray about a window

Some windows carry Ray’s mark, the same symbol that sits at the left of the chat bar. Clicking it tags that window in the bar: a chip appears above the composer and the caret lands in it, so the question you type is about the thing you were just looking at. The chip names the handle Ray’s tools take, so rewrite the abstract or run this on the kidney folder needs no further pointing.

Today the mark appears in:

Every app’s home page, in the header next to the New button: tags the app itself (Data, Commons, Notebooks, Conductor, Studio), so Ray asks which file, dataset, notebook, workflow or document you mean before acting on one.
Studio, in the manuscript toolbar: tags the paper you have open.
Conductor, next to a workflow’s Save and Run buttons: tags that workflow.

Remove a chip with its × before sending if you change your mind.

Attaching things

The plus button in Ray’s box, next to Send, offers three sources:

Upload from this computer — a file that is not in Relay yet.
Choose from Data — anything in your tree: local, Drive, or a connected source.
Choose from Commons — a public dataset asset.

Each attachment appears as a chip above the bar. Selecting notebook cells and then asking a question attaches those cells the same way; the placeholder changes to say so.

Pointing at a file with @

Type @ and part of a name in Ray’s box to list the files in the active Project. Each row shows the file’s name, where it came from (the run that made it or read it) and its role: result, input, intermediate or superseded. Pick one and it is attached as a chip, so Ray gets that exact file and not a similarly named one.

Seeing what Ray is doing

While Ray works, the window, pane or notebook cell it is acting on glows and wears a small tag such as Ray is writing or Ray is running. When Ray opens or focuses something, a pointer labelled Ray travels to it, so you can follow along without reading the transcript.

Asking well

Ray works best with a goal, not a procedure:

Count the nuclei in this image and put the counts in a table.
Open the three videos from cage 4 together and line them up on the TTL pulses.
Write a notebook that loads the h5ad, filters cells under 200 genes, and plots a UMAP coloured by sample.
How do I share only one window? — answered from the docs, with a link.

Questions about Relay itself have a cheaper route: the search box in these docs. Press the Ray mark at its right end (or ⌘↵) and you get a short answer with links to the sections it came from, without spending a Ray run. Continue in Ray under that answer hands the question to the chat bar when you want the full Ray, with your workspace in view. From anywhere in the app, g h or Ask Ray about Relay… in the command palette opens that search.

Follow-ups stay in the same conversation. The New conversation button starts a fresh one; earlier conversations are listed when you focus the bar and can be reopened.

When a run takes a while

Model runs, server pipeline runs and notebook runs keep going after Ray has replied. If Ray started one and the request has a later step (segment, then count), Ray picks that step up by itself when the run ends: a line in the conversation says which run finished and that Ray picked it up. Keep Relay open while you wait, because a reload forgets what Ray was waiting on. Stopping Ray also stops it from carrying on later.

Things worth asking for

Edit a plasmid. With a sequence open, ask swap GFP for mCherry. Ray fetches the real coding sequence from NCBI, replaces the region as a single edit that one undo reverts, and checks the new protein: it starts with M, ends on a stop, has no stop in the middle and stays in frame.
Mark up a slide image. After Ray captures a view onto a Studio slide, ask for a scale bar, an arrow, a box or a label. The marks are drawn into the picture itself, so they show in Studio and in every export, and they are permanent once drawn. A scale bar is sized from the image’s real pixel size; if the capture has none, Ray declines to draw one instead of guessing.
Draw a figure. Pick a model under Draw in the model list (see Choosing a model) and describe the figure, or attach a sketch or photo. For an attached image Ray asks whether to Trace it or Redraw in Relay style unless you already said. Drafts come back as editable SVGs saved to Assets in your Drive, with Copy and Open in SVG editor under them. Click a draft and your next message edits it: make the tumour cell larger.

Approvals and limits

Anything that starts paid compute — a pipeline run on the server, a batch over a folder — asks first. An approval card appears in the conversation with the exact action; nothing runs until you accept it.

Every question counts as one AI prompt against your plan’s daily allowance, which you can check under Settings → Plan & usage.

Choosing a model

The pill in Ray’s box that names the current model appears once a conversation is open. It lists the models in two groups, Chat and Draw:

Model	Good for
Claude Opus 5.5 (default)	Long, tool-heavy work: multi-step analyses, notebooks, anything that touches many files.
Claude Sonnet 5.5	Quick questions and small edits, at a lower cost per run.
GPT-6 Astra	The strongest OpenAI option, comparable to Opus for most tasks.
GPT-5.6 Terra	The fast OpenAI option.
Arrow 2 · QuiverAI	Drawing: describe a figure or attach an image, and it comes back as an editable SVG in Relay style.
Arrow 2 Telos	Drawing busier scenes and full figure panels, at 1.5× the cost.

With a Draw model picked, Ray draws instead of working in the workspace. Switch back to a Chat model for everything else.

The choice sticks for your next runs. Cost per run depends on the model and on how much Ray had to read; each run’s cost is shown in Ray activity.

Where else Ray answers

Slack

Ask in a channel your workspace has connected. Ray acts as you and sees only what you can see.

Your editor or coding agent

Claude Code, Cursor and Claude Desktop reach Relay through the MCP server, with a token you scope.

The command line

Upload, run, watch and stop from a shell, a Makefile, cron or CI.

Studio

A Ray panel inside Studio drafts and edits the document you have open.