ReferenceGlossary

Glossary

A quick-lookup companion for invisible-string's vocabulary — one line per term, with a link to the page that covers it in depth. It's a lookup table, not a tutorial: read the Concepts pages first if you're new to the platform, and come back here when you need a fast reminder of what a term means.

TermMeaning
AgentThe product entity you build: a persona, a model, and the context it's equipped with.
agent editorThe UI surface for building an Agent — its Persona, Model, Context, and Access sections.
agent stepA pipeline step that delegates to a published Agent, dispatched as a child run.
agent versionAn immutable published snapshot of an Agent, identified by a content hash.
allowlistThe workspace's list of provider-and-model pairs Agents are permitted to use.
artifactThe compiled build output (a tarball) of a published agent version — also called a compiled agent.
catalogThe curated, checked-in list of one-click connectors offered in the Add connection dialog.
child runThe agent run an agent step dispatches — a real session, linked from its parent pipeline run's step.
compiled agentSee artifact.
connectionA configured link to an MCP server, at workspace or personal scope.
content hashThe digest of an Agent's identity and definition — persona, resolved model and effort, context, compiler and eve versions, build environment — that identifies its published artifact. An unchanged definition never rebuilds, and two Agents never share an artifact, however alike they look.
contextWhat an Agent is equipped with: MCP connections and skills.
copilotThe AI assistant in both editors that proposes edits as typed, reviewable mutations.
cron expressionThe five-field, UTC-evaluated schedule string that drives a Schedule trigger.
draftThe editable, autosaving version of an Agent or Workflow before it's published.
eveThe open-source agent runtime invisible-string compiles Agents into and runs on.
form triggerA Workflow trigger that renders a shareable field-schema form.
infer stepA pipeline step that makes one direct model call on a workspace preset — text or declared structured output.
ingress tokenThe per-Workflow secret its webhook URL — and its form's submit URL — is authenticated by, shown once at generation.
MCPModel Context Protocol — the standard invisible-string's connections speak to reach external tools.
model presetA workspace-level pointer — provider, model, and reasoning effort — that Agents and infer steps select by name (powerful, balanced, quick).
onComplete replyA Workflow's declared end-of-run Slack reply template — rendered against the final run scope; nothing posts when none is declared.
overlap policyWhether a trigger firing mid-run starts a second run. The default, skip, refuses it — protecting cursor state.
personaThe durable, natural-language description of an Agent's role and working style.
personal scopeA connection or skill visible only to the member who added it, across every workspace they belong to.
pipelineA Workflow's step tree, interpreted by the platform when its trigger fires — no build, no compilation.
probeThe platform's automated health check of an MCP connection.
publishFor an Agent, compiling and building a new immutable version; for a Workflow, validating and snapshotting its pipeline — no build.
reasoning effortHow much a model reasons before answering, from provider-default/none through max.
referenceAn @ token in a step's prompt, arguments, or conditions — @trigger.*, @steps.<slug>.*, @state.*, @item, @now — resolved against the run scope at run time.
registryThe official MCP server registry that invisible-string's community search mirrors and install-time verification checks against.
runOne execution: an agent run within a session, or a pipeline run of a Workflow.
run-as userThe workspace member whose credentials an Agent uses on every dispatch.
sessionA conversation with a published Agent.
skillAuthored markdown-plus-attachments knowledge equipped onto an Agent.
Slack triggerA Workflow trigger fired by mentions or messages in a connected Slack workspace.
stepOne node of a pipeline: tool, infer, or agent does work; for_each, branch, filter, and state control the flow.
step slugA step's handle — the name @steps.<slug> references its output by, unique across the pipeline.
step timelineA pipeline run's record: one entry per executed step instance, with status, attempts, timing, and rendered input/output.
tool stepA pipeline step that makes one deterministic MCP tool call on a workspace connection — no model involved.
tool filterA connection's allow- or block-list narrowing which of its tools an equipped Agent can call.
triggerWhat starts a Workflow: manual, webhook, form, Slack, or schedule.
waitingA run parked on a tool approval, a question, or a session-limit prompt — it still holds the session.
workerThe stateless process pool that boots and runs compiled agents.
WorkflowA standing delegation: a trigger and a pipeline of steps the platform interprets.
workflow stateA Workflow's durable key-value store — written by state steps, read as @state.*, surviving across runs.
workspaceThe multi-tenant boundary every Agent, Workflow, connection, and session belongs to.
world databaseThe per-agent-version Postgres database holding that version's durable execution state.

Easily confused

Agent vs. agent version vs. compiled agent. An Agent is the thing you build and edit — persona, model, and context, changing continuously as a draft. An agent version is one immutable published snapshot of it, identified by a content hash. A compiled agent (or artifact) is the actual eve build output that snapshot produced. You edit the first, runs are pinned to the second, and workers run the third.

A session vs. a run. A session is the whole conversation with a published Agent; an agent run is one execution within it. A session can contain many runs over its life — one per message you send, roughly — and a run finishing doesn't end the session it happened in.

A pipeline run vs. its child runs. A Workflow firing produces one pipeline run — the step timeline. Each agent step inside it produces a child run of its own — a real agent run with a session. The parent holds the ledger; the children hold the transcripts.

A connection vs. the tools it exposes. A connection is the MCP server itself — its URL, its credentials, its health. The tools are what that server advertises once connected, narrowed further by the connection's own tool filter for equipped Agents — while a Workflow's tool step names exactly one of them per call.

Publishing an Agent (which builds) vs. publishing a Workflow (which doesn't). An Agent publish compiles its definition into a real eve project, builds it, and uploads the resulting artifact — a real build pipeline with real failure modes. A Workflow publish only validates the pipeline and snapshots it — it never builds anything, so it's effectively instant.