ARC, inside the project

A conversation per project

Each project keeps its own thread, and several can run at once.

Open ARC inside a project and it is anchored there. Switch projects, browse Notes, go back to the dashboard — the conversation keeps working in the background and is exactly where you left it when you return, still streaming. Because each project holds its own, you can have three analyses running in three projects simultaneously.

Conversation history is stored inside the project itself, so the record of how a result was arrived at sits next to the result.

ARC at Home, across the whole workspace

Step out to the dashboard and the agent widens its scope to every project you have.

From Home, ARC is anchored to the workspace root: it can list and manage to-dos in any project, compare work across projects, and answer questions that span the lot. A conversation you left running inside a project keeps going independently and stays reachable from that project's history.

Background work you can always see and stop

A live pill above the project header for anything running, with a Stop button.

Anything the agent is doing — a chat mid-response, a document review, a background analysis — shows as a persistent pill on the project, whether or not the panel that started it is open. Close the chat panel mid-answer and the pill stays, still stoppable.

There is no hidden background activity in Bench. If something is running, it is on screen.

Plan mode

For anything consequential, ARC writes the plan first and waits for you to approve it.

Rather than beginning a multi-step piece of work and hoping, ARC drafts what it intends to do and shows it to you. You approve, edit the intent, or throw it out. The chat header takes on a distinct tint while a plan is being drafted so the mode is unmistakable.

Approved plans are kept and browsable per project, so you can see what was proposed, what was accepted, and what came of it.

Hand it specific files — including ones outside the project

Drag and drop onto the chat, or use +, to put exactly the right file in front of the agent.

Attachments show as removable chips before you send and stay listed under your message. They can come from anywhere on your machine — a photo on the Desktop, a collaborator's spreadsheet in Downloads — not just the open project.

It reads figures

Point ARC at a plot, gel, microscope image or screenshot and it looks at it.

Images go to the agent directly, including phone photos (HEIC/HEIF), AVIF, BMP and TIFF. Large images are shrunk automatically so the request goes through rather than failing.

Paste a figure from a paper and ask what is wrong with the axis, or hand it your own plot and ask what it would check next.

Standard and Deep

Choose how hard the agent should think for the session.

Standard is for the ordinary majority of work and is quick. Deep is for the problems worth waiting on — study design, an argument you cannot get straight, an analysis with a lot of interacting choices.

When ARC fans work out to sub-agents, they run at the depth you chose: a Deep session gets Deep sub-agents. You can still ask it to use Standard sub-agents for routine legwork.

Long conversations keep going

A session that runs all afternoon compacts itself and carries on.

Extended work no longer dead-ends at a context limit — ARC condenses what has gone before and continues. Follow-up messages in a session also reuse the working context rather than reprocessing the whole thread, so replies start sooner and cost less.

Skills — teach it a procedure once

Write down how your lab does something and ARC follows it every time after.

A skill is a short document describing a procedure: how you normalise a particular assay, the QC gates you insist on, the house figure style, the way your PI wants a methods section written. ARC discovers skills on its own when a task calls for one — you do not have to invoke them.

Skills live as editable files in your workspace or a single project, so there is no rebuild, no vendor lock, and no mystery about what the agent was told.

A public registry of skills

Browse what other labs have written, install it, and ARC can use it immediately.

Published skills live at heurekaskills.com↗ and are browsable from inside the app with search — you never need to know a name or a URL. The browser shows an install count, lists the most-installed first, and puts a medal on the top three. Installs you chose are counted separately from ones ARC made on its own mid-task, so the ordering reflects what researchers pick rather than what the agent happened to reach for.

Ask for something ARC has no procedure for and it will check the registry, install the closest match, and carry on in the same reply. Every skill is reviewed before publication, comes only from the registry, and is checksum-verified on the way in.

Installing never overwrites a skill you already have — if the name collides, the install stops and tells you. Replacing one is a deliberate delete first, so your edits and any files you keep beside a skill are safe.

Publish a skill you wrote

Pass a procedure on, under your own name.

Each skill you authored has a Share button. Before anything leaves your machine Bench shows you exactly which files would become public and how large they are — a pull request on a public registry is readable the moment it opens, and public history is permanent. Pick a category, add up to five tags if you want them, and a maintainer reviews the submission before it goes live.

Shared skills are published under CC BY 4.0↗ and credited to you by name. Your email address is never published, and your local file is left exactly as you wrote it.

Suggest a skill files a request for something that does not exist yet — what you need, a category, and an example source or DOI. Requests are answered on the thread itself, so you can follow one without an account anywhere else.

Custom sub-agents

Define a specialist and ARC delegates to it.

Alongside skills you can define sub-agents with their own brief — a statistics reviewer, a methods writer, a data cleaner. ARC hands the relevant part of a task to them and folds the result back in. Like skills, they are editable files, discovered automatically, with no per-feature wiring.

Research tools

Literature search, reference databases and domain lookups the agent reaches for on its own.

ARC has a set of research tools beyond reading your files: literature search, curated reference databases, and structured lookups it uses to check a claim rather than recall it. A compound written into your record is one that was looked up, not remembered.

A transient connection problem to these tools no longer silently strips them from a message — ARC retries before proceeding, so an answer is never quietly worse than it should have been without telling you.

In Privacy Mode every one of these is switched off and refuses with a clear reason.