Core concepts

A short map of the product’s entities. Full definitions are in the Glossary. Sections marked “for development projects” do not occur in a space without code.

Organisation and members

An organisation is a team’s isolated space. It has an owner and members; one organisation’s data and traffic are not accessible to another. You can create organisations and switch between them.

Development project and repository

• Repository — an imported mirror of your external git repository (upstream).
• Development project — a project that gathers one or more repositories as sources (repo@branch) and in which development takes place.
• Connector — an agent you run in your own network: it connects git that is not reachable from the internet (self-hosted behind a firewall) or spares you from handing us access tokens, and it also runs integrations with trackers and knowledge sources on your side. More in Connector.

Space and knowledge sources

• Space — a workplace without code: questions, deep research and a knowledge base built on the department’s documents. For support, sales, legal, HR, finance, marketing, management, product managers and analysts; it does not turn into a development project. As a scope for permissions and entries it works like a project. See A space without code and the Buff without a repository track.
• Knowledge source — a connection through which the AI agent reads documents where they live: Notion, Confluence, Google Drive, files on a server and network folders, Figma, and optionally the internet. Nothing is copied in advance: only what was needed for the answer is read, and only within the connection’s scope: in the selected sections, or, if nothing is selected, in whatever the account can access (for the internet, empty lists mean the whole public internet, see The web for the model). See Knowledge sources.
• Question (“Ask”) and research — the units of work in a space: an answer in a conversation with links to documents, or a verified document in which a second agent has checked every claim against the sources. See Question or research?.
• The AI agent only reads documents; what it writes is the answer, the research document and knowledge base entries. A system that is not in the connection catalogue can be connected by the company with its own tool.

Integration and connection

The connector is also an integration host, and two different words matter here:

• Integration (plugin) — this is code: a signed bundle from the catalogue that the connector downloads and runs on its side. One integration per system: trackers — Jira (Data Center, Server and Cloud), GitLab, GitHub, Gitea and Forgejo, Bitbucket (Data Center and Cloud — requests in pull requests), Redmine; knowledge sources — Confluence, Notion, Google Drive, Figma, files on disk and network folders; the internet. Besides the catalogue, the connector also runs your own MCP servers.
• Connection — this is the link to a specific system: an address, credentials, a list of repositories or projects, and the development project (or space) where the work goes. Two Jira instances are two connections of one integration, with different secrets and different permissions.

A connection has a slug — a short name derived from its title. It ends up in the names of the tools the model sees, so jira_prod and jira_sandbox read better than two jiras.

A task and its life (for development projects)

A task follows a path: request → refinement/formulation → decomposition into atoms (subtasks) → development in waves → review → publishing. At the key points a person takes over: they approve the plan before the start and approve the result.

The analyst here is a role of the AI agent, not an employee: it turns the request into a specification and asks the author clarifying questions. In a deep research the same role writes the document that a second agent checks.

Atoms and waves (for development projects)

During decomposition a task is split into atoms — the smallest independently buildable units. Atoms are carried out in waves in dependency order; you can see this on the Execution scheme.

Knowledge base

What the organisation knows about itself, apart from the code and from any single task: working rules, facts about the project and pitfalls of the environment. People keep the entries, and the model reads them at every stage and adds to them itself after finished work. This is what outlives the end of a task: without it every next task starts from a blank page. See Knowledge base.

Review and findings (for development projects)

Finished changes go through review: diffs, discussion and findings (from the model and from you) with an “approve” or “changes requested” decision. After approval a pull request is opened in the source repository.