Glossary | nicool.ai Docs | nicool.ai Docs

The vocabulary behind nicoolAI — agents, skills, workflows, context, and the runtime around them — defined in one place.

The words below show up throughout these docs and the product. Definitions are written for how nicoolAI actually uses each term, not the textbook version.

Terms are grouped by theme. The first mention of most of these in any other doc links back here, with the definition on hover — so you rarely have to leave the page you're on to look one up.

Core concepts

Agentagents: A model that operates in a loop rather than answering once: it reads context, decides whether to reply, call a tool, or stop, and repeats until the work is done. nicoolAI is a message-first personal agent that lives in familiar channels and takes bounded action when the boundaries are clear.
Avataravatars · persona · personas: A named, deployable agent persona. An avatar bundles a soul (voice and intent), a set of skills, automation templates, and the access boundaries around them, so the same runtime can present as different specialists to different people.
Context: The full set of information available to the model on a given turn — conversation history, retrieved memory, connector results, and instructions. nicoolAI's runtime works out who is speaking, which conversation this belongs to, and what context should apply before the agent decides how to act.
Context windowcontext windows: The token budget a model can attend to in a single request. It bounds how much history, memory, and tool output fit into one turn, which is why continuity relies on selective memory rather than replaying everything.
Messageinbound message · inbound signal · inbound signals: An inbound signal — a chat message, email, or event — that lands on the webhook path, is verified and persisted, and is processed immediately. If a user reports the agent 'did nothing', the webhook path is the first place to look.
Promptprompts: The instruction and content handed to a model for a single generation. In an agent this is assembled from the system prompt, the running conversation, retrieved context, and tool results — not just what the user typed.
Sessionsessions: A continuous span of interaction that shares state and context. Unlike a one-off assistant session, nicoolAI carries enough continuity between sessions to avoid making a user repeat the important parts.
Skillskills: Not a generic model trick — a defined type of work the agent should do well enough that people can build habits around it (for example a Daily Growth Partner or a Schedule Editor). Skills are role-scoped: some are broadly available, others restricted, and the runtime only reads them through role-scoped loaders.
System promptsystem prompts · system instructions: The persistent instruction block that sets the agent's role, tone, and constraints for every turn, ahead of the user's messages. For nicoolAI this is where restraint and access rules live so behavior stays consistent across a conversation.
Threadthreads: A single ongoing conversation within a channel — a WhatsApp chat, a Slack thread, an email exchange. nicoolAI keeps one thread coherent so the next message builds on the last.
Tokentokens: The atomic unit of text for a model, usually a few characters or a word fragment. Prompts, responses, context windows, pricing, and rate limits are all measured in tokens.
Turnturns: A single pass of the product loop — a message arrives, the runtime resolves context, and the agent replies, takes a bounded action, or escalates. A conversation is a sequence of turns.
Workflowworkflows: An ordered sequence of steps — tool calls, decisions, and hand-offs — that carries a task from a trigger to an outcome. Workflows are how skills compose into something durable rather than a single reply.

Models & inference

Groundinggrounded: Anchoring a response in concrete sources — retrieved documents, connector data, tool results — rather than the model's parametric memory, so answers are decision-ready and checkable.
Hallucinationhallucinations · hallucinate: When a model produces plausible-sounding content that is not grounded in real data. Grounding, retrieval, and restraint — narrowing or escalating when unsure — are the main defenses.
Inference: A single execution of a model against a prompt. In nicoolAI all inference routes through a connected provider, and cost and latency are properties of that call.
Modelmodels · llm · llms: The large language model that turns a prompt into text. nicoolAI is provider-agnostic at the harness level and resolves a sticky per-user model choice, so the same agent can run on different models.
Providerproviders · inference provider: The upstream service that actually runs inference. nicoolAI's inference is bring-your-own: a user connects a provider account and the harness stays provider-agnostic so a second provider can be added beside the first.
Reasoningthinking · chain of thought: The intermediate 'thinking' some models produce before a final answer, spending extra tokens to work through harder problems. Effort can often be tuned per request to trade latency for depth.
Streamingstreamed: Returning tokens as they are produced instead of waiting for the whole response, so replies appear to type out and long generations feel responsive.
Temperature: A parameter that scales output randomness. Lower values make responses more deterministic and focused; higher values make them more varied and creative.

Tools & integrations

Channelchannels: A surface where a user talks to nicoolAI. Channels are where you talk to the agent; connectors are what the agent can touch on your behalf — a distinction worth keeping straight.
Connectorconnectors: An integration that moves the agent from conversation into bounded action inside another system (Google Drive today; GitHub, Obsidian, and personal email planned). A good connector makes one thing obvious: what am I allowing nicoolAI to do?
Function callingfunction call: The model-side protocol for emitting a structured request — a name plus JSON arguments — that the runtime maps to a real tool. It is the plumbing under every tool call.
MCPmodel context protocol: The Model Context Protocol, an open standard for exposing tools and data to agents over a uniform interface. nicoolAI both consumes MCP servers and exposes a narrow one for agent-to-agent use.
MCP servermcp servers: A server that advertises tools to MCP-capable clients. nicoolAI's public MCP surface is intentionally small — OAuth with dynamic client registration and one tool, get_response — so the external contract stays stable while the internal runtime evolves.
OAuth: The standard flow for granting the agent scoped access to an account without sharing a password. nicoolAI uses OAuth for connectors and for its MCP server, including dynamic client registration and PKCE.
Tooltools: A capability the agent can invoke — search the web, read a file, run bash, hit a connector. The model chooses which tool to call and with what arguments; the runtime executes it and feeds the result back into context.
Tool calltool calls · tool use: One concrete request from the model to run a tool with specific arguments. Credentials for a tool call are checked at the moment the call needs them, not assumed up front.
Webhookwebhooks: The HTTP endpoint where inbound signals land to be verified, persisted, and processed. When a channel misbehaves, the webhook path — reachability, auth headers, signatures — is the first thing to check.

Runtime & execution

Croncron run · cron runs · scheduled run: A time-triggered job. Each scheduled execution is its own work scope (cron_run) with its own sandbox, so recurring automations run in isolation from live conversations.
Durable executiondurable · durability: Execution whose progress is persisted so it can pause, recover, and continue across timeouts or restarts. It is what lets orchestration, schedules, and long tool runs be reliable rather than best-effort.
Fan-outfan out · fanning out: Splitting work across several agents or tool calls that run concurrently and merging their outputs. It trades tokens for wall-clock time and broader coverage on research- and review-shaped tasks.
Orchestrationorchestrate · orchestrator: The control layer that decides what runs, in what order, and how results combine — sequencing tool calls, fanning out subagents, and merging their output. It is the 'brain' that keeps a workflow coherent.
Sandboxsandboxes · vercel sandbox: nicoolAI's single execution substrate: everything the agent runs executes inside a Vercel Sandbox, one filesystem per work scope, with roots like /workspace, /schedules, and /gdrive. The model sees paths and files; it does not choose infrastructure.
Snapshotsnapshots: A persisted point-in-time of a sandbox filesystem. If a sandbox times out, Vercel resumes the persistent sandbox from its latest state or snapshot, so liveness is a property of the execution provider rather than application data.
Subagentsubagents · sub-agent · sub-agents: A child agent launched to own a bounded piece of work — a search, a review, a transform — and report a result back. Subagents let one task run as many focused workers instead of one overloaded context.
Work scopework scopes · work unit · work units: The unit nicoolAI keeps one sandbox per — for example user_personal, persona_owner, room_thread, or cron_run. Sandbox names are derived deterministically from the work-unit key, so the same scope reuses the same warm filesystem.

Memory & retrieval

Continuity: The product property of remembering enough to make the next conversation feel coherent, without pretending every situation is equally personal or privileged. It is memory in service of the relationship, applied with restraint.
Embeddingembeddings: A dense vector representation of text whose distance encodes semantic similarity. Embeddings are what make it possible to search memory by meaning rather than exact words.
Memorymemories: The persisted knowledge the agent carries between conversations. nicoolAI keeps enough memory to make the next interaction coherent — not to remember everything forever, but to avoid making a user repeat the important parts.
RAGretrieval-augmented generation · retrieval augmented generation: Retrieval-Augmented Generation: fetch relevant documents at query time and put them in the prompt so the model answers from real evidence rather than memory alone. Retrieval usually rides on embeddings and vector search.
Retrievalretrieve · retrieved: The step of finding and pulling in the content most relevant to a request — from memory, connectors, or a vector store — so it can be added to context before generation.
Vector searchsemantic search · vector store: Nearest-neighbor lookup over embeddings to find the most semantically similar items. nicoolAI's memory approach builds on Convex vector search, favoring verbatim-first, size-adaptive recall.

Access & safety

Bounded actionbounded actions: Doing more than answering — reading a file, sending, syncing — but only when the right connector is available and the boundaries are clear. The point is not that nicoolAI can act; it is that it should only act when the result can be trusted.
Contactcontacts: The canonical identity a sender maps to. Identity is always resolved through contacts.getOrCreateBySender, which preserves the default public role and prevents ad-hoc access drift.
Escalationescalations · escalate · escalates: A clean stop-and-hand-off when context, permission, or certainty runs out. First-line escalation is a skill in its own right: the system should triage and escalate rather than guess.
Fail closedfail-closed · fails closed: The default posture that when access cannot be confirmed the system does less, not more: restricted skills stay hidden, missing permissions block actions, and unclear access yields a denial or a narrower response.
Guardrailguardrails: A required rule that bounds runtime behavior — canonical identity resolution, role-only access decisions, read-only skill access from tools. Guardrails are how safety is enforced structurally rather than left to prompt wording.
Handoffhandoffs · hand-off · hand off: The transfer of a task — with its context — to a human or a different agent when the current one should not continue. Clean handoffs are what make escalation useful rather than a dead end.
Prompt injectionprompt-injection: An attack where content in a message, file, or tool result tries to hijack the agent's instructions. Treating retrieved and tool-sourced text as data rather than commands, plus scoped access, are the defenses.
RBACrole-based access control · role based access control: Role-Based Access Control: capabilities follow from a contact's roles rather than being assumed per conversation. In nicoolAI, restricted skills should never appear by accident and unclear access should produce a denial.
Restraint: A first-class product behavior: a trustworthy agent should narrow, ask, or escalate when it is missing context, permissions, or certainty. Refusal is part of the product, not a failure mode.
Roleroles · contact role · contact roles: An access grant attached to a contact. Runtime access decisions are based on role grants only (contactRoles and skillRoleGrants), never by querying skills directly, which keeps a default public role from drifting into more access.

Platform

Codex: An upstream coding agent nicoolAI integrates with, including a connected-sub inference path (bring-your-own tokens) and PR review. Inference can be hard-locked to route through a connected Codex sub across channels.
Convex: The backend platform behind nicoolAI — the durable, reactive layer for data, orchestration, and the LLM loop. In the brain-and-hands split, Convex is the brain and the sandbox is the hands.
Evalevals · evaluation: A structured check that runs the agent against known inputs and scores the output, so changes to prompts, tools, or models can be measured rather than eyeballed.
httpActionhttp action · http actions: A Convex function that handles a raw HTTP request, used for things like the web chat turn. It is the entry point where auth (a JWT in the request) can be verified before running privileged work.
Materializationmaterialize · materialized: The step that instantiates a pack's contents as real, tenant-owned state. It is provenance-tagged so later installs and upgrades can tell what came from which pack.
Model selectormodel picker: A per-user, sticky preference for which model handles their requests, resolved by the harness before inference. It lets different users run the same agent on different models.
Packpacks · avatar pack · avatar packs: An installable unit — soul, skills, automation templates, and a qualify interview — that materializes into a tenant to stand up an avatar. Materialization is provenance-tagged so packs can be installed later and upgraded cleanly.
Provenanceprovenance-tagged · provenance tag: Metadata that tags materialized records with the pack and version they came from, so install-later and upgrade flows can reconcile changes without clobbering user edits.
Qualify interviewqualify interviews: An agent-led interview shipped inside a pack that gathers what it needs to configure the avatar for a specific person or team during onboarding.
Room threadroom threads · room_thread: A shared conversation with more than one participant, treated as its own work scope (room_thread) with its own sandbox and continuity, distinct from a user's personal DM space.
Soulsouls: The part of an avatar pack that defines who the persona is — its voice, priorities, and disposition — separate from the skills and automations layered on top.
Tenanttenants · multi-tenant · multi-tenancy: A customer boundary that owns its own data, avatars, and access. nicoolAI is multi-tenant: one deployment serves many isolated tenants, which is why access is scoped rather than assumed.