Reference

Glossary

Every term that appears in the docs with a load-bearing definition. Skim this once before reading the rest of the reference section and the platform's vocabulary will feel familiar everywhere.

Updated today

Platform terms

The names that show up across the surface.

Ultron: The platform itself. A multi-persona, skill-based chat product running on a containerized cloud runtime with a Supabase backend and a small fleet of scheduled workers running alongside the app.
Workspace: One user's slice of the platform. Every row that belongs to a user carries user_id and is RLS-scoped to that user. A workspace is the unit of billing.
Project: A user-defined grouping of conversations inside a workspace. Optional. Adds project-level instructions and project-scoped memory.
Conversation: A chat thread. One row in conversations. Belongs to a user and optionally a project; carries the active skill and the active mission.
Run: A single end-to-end execution of a worker pipeline (a carousel render, a scraper crawl, a fundraising worker job). Tracked in a *_runs or *_jobs table.

Chat loop

The vocabulary of one assistant turn.

Turn: One full user-to-assistant exchange. Begins when the user sends a message; ends when the assistant either streams a final response or pauses for approval.
Chat run: One row in chat_runs. Tracks the lifecycle of a turn: started_at, completed_at, model, status, error if any.
Tool event: An event emitted while a tool runs (start, partial output, end). Streamed over SSE and stored on messages.tool_events for replay.
Pause point: A spot in a turn where a tool needs human approval before it can run. Stored in pause_points and resumed on user decision.
SSE: Server-Sent Events. The transport for streaming assistant tokens and tool events to the browser. One long HTTP response with newline-delimited events.
Channel: A label on a message that says where it should render (default, canvas, share). Used by the renderer to route content to the right surface.

Skills and personas

How the platform organises capabilities.

Skill: A bundle of prompt, tools, and rendering rules that the agent loads to do one job. 57 skills ship today across 8 categories. A skill is a delegated sub-agent: run_skill spins up its own agent loop (own model tier, own scoped tools, own turn budget) and returns a finished deliverable.
SKILL.md: The markdown file that defines a skill. Contains the prompt, the tool allowlist, the rendering rules, and example transcripts.
Recipe: A vetted build/deploy playbook the model reads and follows itself on a shell turn — no sub-agent, no separate model call. 7 ship today; the model pulls a recipe's full body on demand via load_recipe when a shell and repo are bound. Distinct from a skill: a skill runs on its own and hands back a result, a recipe is a procedure the model follows itself while building.
Capability: A high-level action a skill exposes (gmaps-leads, send-email, generate-carousel). The platform translates capabilities into concrete tool calls.
Persona: A pre-set agent identity that bundles a default skill set and a system tone (Cortex, Specter, Striker, Pulse, Sentinel, Amplify, Counsel).
Tool: A function the agent can call. 91 native tools plus 7 from the billing integration plus 1 human-in-the-loop tool = 99 callable functions.
Tool catalog: The static registry exported from chat-tools.ts. Maps a tool name to its schema, its handler, and its UI renderer.
HITL: Human in the loop. The single tool that asks the user a question mid-turn and waits for the answer before continuing.

Memory

The three storage surfaces and the compression cascade.

Agent memory: Short structured facts the agent keeps across turns. Stored in agent_memories with a 768-dim embedding for similarity lookup.
Brain file: A file the user uploaded to the workspace. Stored as binary in Supabase Storage; chunks indexed in file_embeddings with HNSW.
Hybrid retrieval: The query strategy on brain files. Combines pgvector cosine similarity with a tsvector full-text rank, then re-ranks by recency.
Cascade compression: The four-layer summarisation that keeps long conversations inside the chat-model context window. Recent verbatim → recent summary → older summary → distilled facts.
Note: A free-form sticky note the agent can write into and read from. Stored in notes; not embedded; addressed by title.

Canvas and workspace

Surfaces other than chat.

Canvas: A non-chat rendering surface for structured outputs (tables, kanban, dashboards, slide decks). Triggered by the chat button or a /techniques flow.
Block: One renderable unit on a canvas. 30+ block types live in the CanvasBlock registry; each takes a typed data prop.
Technique: A pre-defined canvas-producing workflow. Listed at /techniques. The chat surface can invoke any technique without the user typing the underlying prompt.
Workspace sandbox: An isolated per-conversation sandbox on managed Kubernetes that gives the agent a real Linux desktop, reachable from the browser as a live remote desktop view. One per active user.
Sandbox snapshot: A point-in-time image of a sandbox. Used to resume the same desktop state on the next session instead of cold-booting.

Ultron Workers

The internal name for the platform's scheduled workers and the scraping actors.

Ultron Worker: A scheduled worker running alongside the app, deployed under the platform's account. Six exist: cron, marketing-swarm, model-shim, scheduler, fundraising, smtp-verify.
Marketing pipeline: The seven-stage carousel generation flow inside the marketing-swarm worker: Retrieve → Plan → Validate → Render → QA → Promote → Respond.
Spec: The structured CarouselSpec produced by the planner and consumed by the renderer. Strictly freeform-only slides.
Exemplar: A past carousel run that scored at or above the promote threshold. Stored as PNGs plus spec JSON in R2; brief embedded in Vectorize.
Star: User-driven promotion. Forces a run into the exemplar index even when its score was below the auto-promote threshold.

Scraper Workers

The other half of the worker fleet.

Scraper Worker: A scraping actor deployed under the platform's name on the scraping platform. Six in production: gmaps, jobs, email finder, linkedin, twitter angels, plus the leadgen and research orchestrators.
Webhook flow: The pattern where every scraper posts results back to /api/apify-webhook with a source query param. The platform's webhook handler dispatches per source.
Residential proxy: An IP pool sourced from real end-user devices. Used for Google Maps because datacentre IPs are blocked aggressively.
li_at: The session cookie LinkedIn uses for authentication. Required by the jobs and linkedin scrapers; lives only in the caller-provided actor input.
Dedup key: The column or expression each scraper uses to identify duplicate rows on upsert (place_id for gmaps, jd_url for jobs, profile_id for linkedin).

Infrastructure

Platform terms a reader sees in deployment and reference contexts.

RLS: Row-level security. Postgres feature that filters rows on read and write based on a policy expression. Enabled on 110 of 113 public tables.
pgvector: Postgres extension that adds the vector type and HNSW indexing. Used for brain-file retrieval and agent-memory similarity.
HNSW: Hierarchical Navigable Small World. The graph-based index pgvector uses for approximate nearest-neighbour search over embeddings.
MCP: Model Context Protocol. The protocol the chat surface uses to talk to external tool servers like a puppeteer MCP or a private vendor MCP.
AEO: Answer-Engine Optimisation. The doc-writing style that biases toward TL;DRs, At-a-glance lists, glossaries, and short crisp paragraphs that LLMs can quote cleanly.

Billing

The vocabulary of the credit ledger.

Plan: A subscription tier (Starter, Max, Ultra) priced as a recurring Stripe price. Each plan has its own monthly credit grant and feature flags.
Credit: The platform's billing unit. One credit roughly maps to one mid-sized model call; cheap calls cost fractions, expensive ones cost multiples.
Credit grant: A non-expiring or expiring pool of credits given to a workspace (plan inclusion, top-up purchase, manual grant).
Credit ledger: credit_transactions, append-only. Every debit and credit lives here with a description and a source reference.
Suppression: An email address that should never receive outbound mail. Captured from bounces, complaints, and manual unsubscribes.