The words of the product

Get started

Seven words carry the whole product. Two of them name more than one thing. Read this page before the rest.

Project

A project is one activity: one site, one database, one repository, one workspace. It is also the unit of billing, of roles and of quotas. The free plan allows one active project; Pro and Max allow as many as you want.

A project holds two threads, a file tree, its versions, its collaborators, its agents and its automations. Everything in the product happens inside a project. The address /projects is not a list: it reopens the last project you had open.

Mode

The word names two things, in two places.

The thread mode is the switch above the message box: Code or Automations. It decides which conversation you are in, and each thread keeps its own memory of the exchange.

The response mode is the selector under the message box, in the Code thread only: Ask, Plan, Build, Agent. It decides what the assistant may do with your message. The four modes gives the full table, what each one may call, and how the badge behaves before you choose.

Agent

Two unrelated things carry this name. One is the response mode above, covered in The four modes. The other is a role you write in the Automations face: a title, a trigger, a business function, a job description, a set of tools, an autonomy level and a cadence.

An agent's business function and tools are written at creation and never again. Its title, trigger, job description, cadence, autonomy and model stay editable. Deleting an agent deletes its run history with it.

Automation

An automation is a recurring instruction handed to one agent. What separates it from an agent's own cadence is a single field, the instructions, which is why one agent can carry several automations.

Its cadence is once only, every day, or certain days of the week. It becomes active as soon as it is created. Editing its instructions, its schedule or its model freezes a new version, and a queued run always executes the version that left, never today's. Archiving it stops it from leaving again and keeps its run history.

Memory

Memory is what your agents re read before every job: a constraint, a tone, a name that matters. It has exactly two scopes. Business rules serve every agent of the project. Agent rules serve the one agent they name. No agent ever reads another agent's rules.

Rules you wrote and rules MaShop inferred from the work stay in separate groups, and an inferred rule is labelled as inferred when it reaches the model. Correcting an inferred rule turns it into one you wrote.

Two other memories exist and are easy to confuse with this one. The chat keeps the last 20 messages of the current thread, which is not the Memory screen. And the working context an assistant reuses between messages is dropped after 24 hours of inactivity.

Connector

A connector is one of your outside accounts, attached to the project so agents can act in it. Six are official: GitHub, Supabase, Netlify, Vercel, Stripe, Resend. They bring 23 tools between them, each classified by hand as read, write or critical. Four are critical: pushing a commit, executing SQL, changing DNS, creating a refund.

Each tool has three possible states: allow, ask, or turn off. With nothing set, read tools are allowed and everything else asks. Allow wins over an agent's autonomy, so a tool set to allow runs even for an agent that normally prepares and waits. Only the project owner attaches, revokes or changes a permission.

Any other service reachable over MCP can be added too. Its tools are relayed to that server and run there, under the same permission gate. Those servers are published by their authors and verified by nobody, and MaShop says so on the screen that offers them.

Credit

One credit is one cent. The cost of a generation is computed from the real provider cost, multiplied by the platform margin, converted at 100 credits per dollar, with a floor of 0.01 credit. The multiplier is uniform: no model carries its own markup, and the multiplier printed next to a model name in the selector is decorative.

Credits are drawn in a fixed order: weekly credits first, then subscription credits, then your wallet of top ups and bonuses. The balance shown at the top of the screen is the sum of the three. Subscription credits are reset and re granted at each renewal, so what you did not spend is lost. Purchased credits expire one year after purchase.

Screens that read the same for everyone

A few screens carry their text in the code rather than in the translations, so your language setting does not reach them. Four are French: the share window, the project creation window, the invitation landing page and the transfer error page. The onboarding wizard is English for the same reason.

Some screens also still carry older names for things this page has renamed. The glossary lists them and says what each one means today.