Fundamental AI terminology in Sculptor — provider, model, prompt, skills, MCP, and agents

Teams adopt Sculptor faster when everyone shares the same vocabulary. This article defines the core AI terms you see in Settings, workspace Chat, and Agentic Strategy Pack runs—and how they map to product behaviour (not generic hype).

AI provider

An AI provider (sometimes “LLM provider”) is the service or runtime that actually runs the model and returns completions.

In Sculptor you pick the active provider in Settings (tabs such as Ollama (local), Ollama (cloud), OpenRouter, OpenAI, Anthropic, Google Gemini). Organisation workspaces also offer Organisation, Personal, and Conversation credential scopes in the composer—see Why three LLM layers. That choice decides:

• Where requests go — cloud API, same-origin proxy, or your machine (http://127.0.0.1:11434 for local Ollama).
• How you authenticate — API keys, bearer tokens, or org-level credentials on Company account.
• Which model list you can pick — e.g. List cloud models on the Ollama Cloud tab.

Provider is not the same as Sculptor. Sculptor is the coaching app, library, and orchestration layer. The provider is the engine you configure. Organisation workspaces can prefer company credentials; individuals often use bring-your-own-key stored encoded in the browser.

For a hands-on provider walkthrough, see How to set up Ollama Cloud in Sculptor.

Model

A model is a specific trained checkpoint (identified by a string id) on a provider—gpt-4o, claude-sonnet-4-20250514, llama3.3, or whatever your Ollama/OpenRouter account exposes.

Layer	Question it answers
Provider	Who hosts the API?
Model	Which brain on that API?

Models differ in reasoning quality, context window, speed, cost, and tool-calling support. Sculptor honours the model you select in Settings (and org defaults where configured) for Chat, Agentic phases, and MCP agent turns—agents should not silently ignore your configured backend.

Practical tip: start with one stable model for strategy work, then experiment. Switching model without switching provider is common; switching provider usually means new keys and a different model picker.

Prompt

A prompt is everything the system sends to the model to shape the next reply—not only the sentence you type in the composer.

In Sculptor a single assistant turn typically assembles:

1. System prompt — Strategy Pack specialist rules (orchestrator, /intake, /market, etc.), journey context, optional B2B playbook, and enabled AI Skills.
2. Conversation history — prior user and assistant messages in the thread.
3. Your message — the text (and attachments) you send now; may include slash commands like /brain or /finance.
4. Grounding — library excerpts, pinned assets, or Agentic prior-phase excerpts when relevant.
5. Tool results (when MCP is on) — outputs from filesystem, GitHub, Context7, org HTTP MCP, or Sculptor’s own tools—wrapped so the model treats them as untrusted external data.

What you type is the user part of the prompt. What you do not see in the UI is still prompt: specialist methodology, skills catalog, and orchestration directives in Agentic mode (for example instructions to avoid “reply to continue” questionnaires between phases).

Good prompts in Sculptor are specific about stage, audience, and constraints—“B2B SaaS, EU, pre-PMF, challenge our pricing story”—rather than “be smart.” Slash commands and the Commands palette route you to the right specialist so you do not paste ten pages of methodology by hand.

Skill

In Sculptor, an AI Skill is a named block of custom instructions you maintain from the Skills pill in the chat composer (beside Commands).

• Each skill has a name, instructions, and an enabled toggle.
• Only enabled skills affect the coach.
• Skills are stored in this browser only (localStorage) in the current product version—they are your personal “house rules,” not org-wide database presets yet.

On each reply, enabled skills are merged into the system side of the prompt. With Settings → Enable MCP for assistant replies, the model can also call sculpt_read_ai_skill to load full skill text on demand.

Skills complement Strategy Pack phases; they do not replace /intake or /validate. Prefer pack methodology for process, skills for tone, taboos, and standing facts.

Deep dive: AI Skills in Sculptor.

MCP (Model Context Protocol)

MCP is an open protocol for connecting AI clients to tool servers—small programs or HTTP gateways that expose actions like read_file, search_docs, or list_issues.

Think of it as a standard plug between the assistant and the rest of your stack:

• MCP server — advertises tools and runs them when asked.
• MCP client — lists tools, lets the model choose one, executes it, returns JSON/text into the chat loop.

Sculptor participates in MCP in two directions:

1. Consume external MCP — Settings and org Company account → MCP servers wire filesystem, GitHub, Context7, or custom HTTP MCP into POST /api/chat/agent when MCP is enabled for replies. See Connect MCP servers to Sculptor.
2. Expose Sculptor as MCP — Claude Desktop, Cursor, or other clients can call library, conversation, and agent-run tools over https://sculptor.scrumdesk.com/api/mcp/http. See Use Sculptor as an MCP server.

MCP is optional. Without it, the model only sees text you provided. With it, the model can act under your credentials and configuration—still subject to membership checks and org privacy toggles.

AI agent

In the wider industry, AI agent often means any system that plans, calls tools, and loops until a task is done—not just a single chat completion.

In Sculptor, use these distinctions:

Idea	What it usually means here
MCP agent turn	One assistant reply path using POST /api/chat/agent: the server may run multiple tool rounds (MCP + Sculptor tools) before streaming final text to the UI. Enabled when Enable MCP for assistant replies is on in Chat.
Agentic (Strategy Pack) mode	A workspace mode (header toggle): a multi-phase run that walks slash-command phases in order, writes markdown library assets, and can chain dozens of specialists without you answering between steps.
Specialist “agent” prompts	Markdown files under lib/prompts/strategy-pack/prompts/ (e.g. market-intel, finance)—personas for chat, not a separate product login.

So “agent” in a product demo might mean tool loop; in Sculptor marketing we often mean Agentic runs—autonomous phase execution with durable outputs.

How AI Agent differs from AI Chat in Sculptor

Chat and Agentic are two explicit modes in the workspace. They are not two names for the same screen.

Chat (user-driven coaching)

• You send each message when ready; the thread waits for you between turns.
• Replies stream token-by-token (default path via provider proxies).
• You pick specialists with slash commands or the command palette (/brain, /brand, …).
• Optional MCP per turn when enabled—same tool loop as above, but you decide when to continue.
• Best for exploration, debate, one domain at a time, and refining a section after you have read the output.

Agentic (orchestrated Strategy Pack run)

• You start a run (brief, depth toggles, optional library assets); the orchestrator advances phases without expecting answers between steps.
• Each phase uses the same specialist system prompts as chat, plus synthetic user messages and capped prior-phase context so later phases build on earlier artefacts.
• Outputs are saved as library markdown assets per phase; you get a combined document when the run completes.
• Non-interactive by design in v1: models are steered not to end with “please answer these twelve questions to continue.”
• Best for breadth—running many phases in journey order when context is thin but you want a full pack draft to edit.

Side-by-side summary

Dimension	Chat	Agentic
Who paces the work?	You	Orchestrator
Between phases	You choose when to run /market or stop	Phases chain automatically
Typical output	Messages in one conversation	One asset per phase + combined view
MCP	Optional per assistant turn	Same stack possible; streaming proxy is default for phases
LLM settings	Your provider + model from Settings	Same provider stack for phase calls
When to use	Precision, Q&A, iteration	Speed across the pack, first draft library

Both modes share Strategy Pack methodology, library-first artefacts, and respect for configured provider and model. They differ in control and packaging: conversation versus pipeline.

For the product story from chat to library, see From specialist chat to a library-ready Strategy Pack. For phase list and journey order, see Strategy Pack — 35 phases guide.

How the pieces fit together

1. Configure provider and model in Settings.
2. Choose Chat or Agentic for how work is paced.
3. Rely on prompt assembly (pack + your message + skills + library).
4. Turn on MCP when the coach must read repos, docs, or external systems.
5. Use skills so tone and constraints survive every thread.

What to try next

1. Open Sculptor, confirm Settings → active provider and model, send a one-line test in Chat.
2. Set Reader persona and Coach responses on the project—see Coach persona and answer style.
3. Enable one Skill with your company briefing; notice how replies stay consistent across threads.
4. If you use tools, enable MCP for assistant replies and validate org servers on Company account.
5. When you need a wide first draft, start one Agentic run; return to Chat to stress-test weak sections.

The glossary above is meant to stay stable as features evolve—check What's new for release bullets and this blog for the longer “why.”