Concepts

The words the app, the API and the plans use, each in one paragraph.

Workflow

A named, versioned recipe of steps across the family products, with its inputs, the products it needs, its approval policy and its trigger. Each use case on this site is one workflow. Its plan is an OpenSpec change.

Run

One execution of a workflow, with an id like R-2041. A run has a status (queued, running, waiting for approval, done, failed, canceled), a step timeline and, when it ends, a signed record.

Step

One unit of work inside a run. A step calls one product, through its MCP tools or API, as the run's agent. Each step reports what it saw with an evidence state: measured, inferred or missing.

Approval

A gate before a step that touches the world. The runner asks PromptCompute (request_approval) with the exact action and arguments. A person decides, and the step runs only with a verified resume credential and the final arguments.

Record

What a run leaves behind: the task id, the append-only list of events, each approval's signature and a hash over the whole. Records can be published to AI Source of Truth so anyone can check them later.

Brain

The shared memory: lessons, decisions and past errors agents read before they act. A person publishes and can revoke each lesson. Steps read it with search_brain and file new lessons with propose_brain_entry.

Workflow request

One run of a workflow. It is the unit your plan counts: 100 per user per month on the $25 plan, 500 on the $50 plan, pooled across the org. Proposed counting rules: a run counts when its first step starts; retries and resumes inside the run do not count again; a dry run does not count.

Agent token

A short-lived token from PromptReports ID that names the person (sub), the agent and its version (act) and the org. The runner trades its own grant for one token per product (RFC 8693 token exchange).

Connection

A product linked to your org, with its key or token and status. Each connection is limited to that product's own actions, so a leaked key cannot act elsewhere.

Evidence state

Each fact a step reports is measured (read from the source), inferred (worked out, with how sure), or missing (not available). A run never turns missing into measured.

Model key

Every model call a run makes will use your org's own key with your model provider. Your plan pays for workflow requests, not for model calls.

Family assistant

One chat that will answer questions across every product and act on them, each change going to a person as a PromptCompute approval. It answers what the Ask across the family use case asked. Planned: it will live in the app and as an ask tool on the family MCP server.