One line per term, linked where a page goes deeper.
Actor — who an operation runs as: you, one of your automations, or a contracted outside party such as a connected agent or a customer. Every audit entry names its actor.
Approval — a held operation waiting for your decision, durable until answered, often with editable arguments. See Grants and approvals.
Archive — the single encrypted .recued.archive file that carries a full
server backup. See Backup and restore.
Audit — the local record every execution writes: actor, channel, reads, writes, outcome. See Audit and memory.
Batch mode — one model call applying a contracted AI function to every
element of an array, switched on by llm.id_field.
Browser Bridge — the browser extension that observes declared DOM triggers, performs page reads and actions, and raises OS notifications on the server's behalf. Runs only signed commands; holds no data.
BYOK — bring your own key: your paid or self-hosted model endpoints in the two generative AI slots. See Set up AI.
Circuit breaker — the guard that disables a reactive recipe after repeated consecutive failures and notifies you; re-enabling is manual.
Claim link — the mailed, single-purpose link a customer follows to collect an access token — the token itself is never mailed. See Selling access.
Connection — a named, reusable record for an outbound endpoint (API, MCP server, or notification target); credentials stay on the server. See Connections.
Contract — the stored grant record the enforcement boundary consults: what an installed capability, session, or customer tier may do.
Customer — an outside party a seller grants tiered access to; connects with any MCP or OpenAI-compatible client, no Recued required.
Door — a deliberately opened, externally reachable channel — MCP, reception, or messenger — carrying its own per-tool grants and credential.
Enforcement boundary — the single gate every operation passes: resolve actor and capability, check grant and risk tier, execute locally, write audit.
Embeddings slot — the dedicated provider and model for vector-producing recipes and housekeeping. It is never used by chat or normal generative AI and has no fallback to the free pool or BYOK slots.
Enrichment — a precomputed fact attached to a warehouse record (a score, a signal), produced deterministically or by AI under per-topic trust.
Free pool — your own free-tier model API keys, used before paid keys when routing AI calls.
Grant — permission given in advance: which operations a pack, door, or session may use, at what access level, for which audience.
Housekeeping — idle-driven server maintenance: compaction, cache eviction, enrichment refresh, CRM reconciliation.
Ingredient — a manifest declaring one operation — an HTTP call, page action, AI function, storage or connection call — with its inputs, outputs, and risk tier. See Ingredient manifest.
Kitchen — where recipes and ingredients are authored, forked, and tested before publishing.
LLM gateway — the server's OpenAI-compatible chat endpoint
(/v1/models, /v1/chat/completions), letting any base-URL-plus-key chat
client hold governed turns with a granted contract. See
Connect to Recued.
Marketplace — the public catalog of recipes, ingredients, and packs; free to browse and install, publishing needs a free account.
MCP — the open protocol external AI agents use to call your server's tools. See AI chat and MCP.
Memory — audit entries graduated into a linked collection with provenance, powering per-entity timelines.
Pack — the installable bundle of recipes, ingredients, and policy; the unit grants attach to. See Recipes, ingredients, and packs.
Pairing — attaching a client (webclient or Bridge) to your server with a narrow, revocable token; server-direct, never through the cloud.
Pro — the optional $5.99 USD/mo convenience bundle: a <handle>.recued.net
name plus an auto-renewed certificate. Gates no capability.
Publisher handle — your free marketplace namespace; everything you publish lives under it.
Reception — your server's public front door: pages, forms, scheduling links, and file drops, with incoming requests held for review. See Reception.
Recipe — a readable JSON description of a unit of work: steps over ingredients and transforms. See Authoring recipes.
Recovery key — the 24-word phrase that authorizes pairing and unlocks the data your server encrypts, including backup archives.
Server — the self-hosted engine and root of trust: warehouse, credentials, scheduler, audit authority, and the only thing that touches your accounts.
Tier — a seller's named access level, backed by a customer contract template that fixes its grants and usage policy.
Timeline — the per-entity chronological feed compiled from provenance links across mail, calendar, CRM, annotations, and memory.
Transform — a pure, deterministic, free function recipes compute with. See Transforms.
Trigger — the URL patterns a recipe is designed for (auto-run), as
distinct from trigger_steps, the self-gating phase of a reactive recipe.
Warehouse — the local store of mail, calendar, contacts, files, and CRM snapshots that grounds recipes and AI answers. See The warehouse.
Watcher — a reactive recipe's trigger ingredient observing mail, files, calendar, time, or pages, deciding whether a tick proceeds.
Webclient — the thin app at app.recued.com (also served by your own
server): display and input, no engine, no credentials.