Reference

Glossary

Plain-English definitions for the terms you meet across the Atlas API, auth model, platform, and AI surfaces. Filter to find one fast, or link straight to a term.

43 terms

API and requests

The words you meet the moment you make a call against the public REST API.

Personal Access Token (PAT)atlas_pat_...: A long-lived credential you mint in your account settings to call the API as yourself. It carries your own permissions and the scopes you grant it, and you can revoke or rotate it at any time. See authentication.
Bearer tokenAuthorization: Bearer ...: Any token you send in the Authorization header to prove who you are. A PAT or an OAuth access token both travel to Atlas as bearer tokens.
OAuth scopetasks:read: A named permission such as tasks:read that bounds what a token can do. A request only succeeds if its token carries every scope the endpoint requires.
Scope minimisation: Granting a token the narrowest set of scopes it needs and no more. It limits the blast radius if a token ever leaks, so start read-only and widen only when a call needs it.
Idempotency keyIdempotency-Key: A unique value you attach to a write request so retrying it never creates a duplicate. Atlas remembers the first result and replays it for any repeat that reuses the same key. See idempotency.
Rate limit429: A cap on how many requests you can make in a window. Exceed it and you get a 429 with a Retry-After hint telling you when to try again.
Rate-limit classread | write | ai: The bucket an endpoint counts against. Reads, writes, and AI calls have separate budgets so a burst of one never starves the others. See rate limits.
Problem Details (RFC 9457)application/problem+json: The standard error shape Atlas returns, with type, title, status, and detail fields. One predictable body for every failure. See errors.
Cursor paginationcursor: Paging by an opaque cursor instead of a page number. You follow the next cursor from each response until it runs out, which stays correct even as data changes. See pagination.
ETagETag: A version fingerprint returned with a resource. Send it back later to ask whether the resource changed, and to guard your writes against a stale copy.
If-MatchIf-Match: A request header carrying an ETag. The write only lands if the resource still matches, otherwise you get a 412 and know someone changed it first.
Optimistic concurrency: The pattern behind ETag and If-Match: let anyone read, and reject a write only when someone else changed the record first. No locks, and no silent lost updates.
Webhook: An HTTP callback Atlas sends to your URL when an event happens, so you react to changes without polling. See webhooks.
Signed webhook: A webhook whose body is signed so you can verify it truly came from Atlas and was not tampered with. Check the signature header before you trust the payload.
Server-sent events (SSE)text/event-stream: A one-way stream that pushes events over a single long-lived HTTP response. Atlas uses it to stream AI output and live updates as they happen.

Auth and access

How people and machines sign in, and how you keep the right ones out.

Tenant: One isolated customer account. Every workspace, user, and record belongs to exactly one tenant, and nothing crosses the boundary.
Super admin: An internal Atlas operator role that can administer tenants for support. It is separate from any customer's own admins and is closely audited.
Single sign-on (SSO): Signing in to Atlas through your company identity provider instead of a separate password, so access follows your directory.
SAML: The enterprise SSO protocol Atlas supports for federated login. Your identity provider asserts who a user is, and Atlas trusts that assertion.
SCIM: The standard for provisioning and de-provisioning users automatically from your directory, so joiners and leavers sync without manual work.
Bring your own key (BYOK): Supplying your own encryption or model-provider key so Atlas uses your key material rather than its default. You stay in control of the key and can rotate it.
Two-factor authentication (2FA): A second proof of identity beyond your password, such as a code from an authenticator app. Admins can require it across a tenant.

Platform model

The objects Atlas is built from, and how plans decide what you can reach.

Workspace: A shared container for a team's projects, tasks, and data inside a tenant. Most day-to-day work and most API calls are scoped to a workspace.
Project: A unit of work that groups related tasks toward a goal, with its own members, timeline, and status.
Task: The atomic unit of work in Atlas: something to be done, with an assignee, a state, and a due date. Most of the API and most MCP tools revolve around tasks.
Entitlement: A capability a plan unlocks, such as a module or a feature flag. Atlas checks entitlements on the server, so gating fails closed rather than leaking a paid feature.
Plan gating: Turning features on or off based on the tenant's plan and its entitlements. A gated feature returns a clear error rather than half-working.
Sandbox tenant: A throwaway tenant for testing integrations safely, isolated from your real data. Break things here without consequences before you touch production.
Governance: The controls that keep Atlas compliant and auditable: audit logs, approvals, data retention, and access reviews.

AI and agents

The vocabulary of driving Atlas with AI, from the protocol down to a single run.

Model Context Protocol (MCP): An open standard that lets AI agents work with Atlas through a typed set of tools, resources, and prompts. Point Claude, Cursor, or Cline at Atlas and drive it in plain language. See MCP for AI agents.
MCP tool: An action an agent can invoke over MCP, such as creating a task. Each tool has a typed schema so the model knows exactly what to send.
MCP resource: A piece of read-only context an agent can fetch over MCP, such as a task or a project, addressed by a stable URI.
MCP prompt: A reusable, parameterised prompt template the server offers to agents, so common workflows start from a vetted place instead of a blank box.
Agent: An AI actor that works inside Atlas on your behalf, reading context and taking actions through tools.
Agent run: A single execution of an agent from start to finish, recorded so you can review what it did and why.
Approval: A checkpoint where a human must confirm before an agent's proposed action takes effect. It keeps a person in the loop for anything sensitive.

Connect and operate

How Atlas talks to your other tools, and how you know it is healthy.

Connector: A packaged link to an outside product such as Slack or Salesforce, with the credentials and scopes Atlas needs to talk to it. See connectors.
Integration: A configured, live connection built on a connector, including how objects map and sync between the two systems. See integrations.
Changelog / release: The running record of what shipped and when. Each release is a dated, versioned entry you can read or subscribe to.
Status page: The public page that reports whether Atlas is healthy right now, plus any active incidents or planned maintenance.
Incident: A logged period of degraded or failed service, with updates posted as the team investigates and resolves it.
Maintenance window: A scheduled, announced period when Atlas may be briefly disrupted for planned work.
Uptime: The share of time a service was available over a period, usually shown as a percentage on the status page.