Concepts

AI

How Monospace's AI works — the embedded assistant and the MCP server, and the tools and permissions they share.

Overview

Monospace exposes AI in two ways, both built on the same foundation:

The Assistant — an AI chat panel inside the Studio.
The MCP server — a Model Context Protocol endpoint for external AI agents and tools.

Both work through the same set of tools and run under the permissions of whoever — or whatever — is using them. Neither can do anything the acting user or API key is not already allowed to do.

What the AI Can Do

Through either surface, the AI works with your workspace using the same capabilities:

Read the schema — collections, fields, and relations.
Query items, using the same filtering the API supports.
Create, update, and delete items.
Read and change the schema.
List the workspace's data sources.

Every action is checked against the acting identity's permissions: item operations follow collection permissions, reading or changing the schema requires the data-model entitlement, and listing data sources requires both dataModel:read and dataSource:read. The AI cannot do more than its user or key already can.

The Assistant

The Assistant is an AI chat panel in the Studio. Ask it to find, summarize, or change data in natural language; it keeps context within a conversation, and you can start a new chat at any time. Because it runs as you, it sees and changes only what your roles allow.

Configuring the Assistant

Configure the Assistant per workspace under Workspace Settings > General > AI. Connect a provider with its API key, then choose the models the Assistant uses.

Setting	Controls
Connector	The model provider. Monospace supports OpenAI.
API key	The provider credential. It is encrypted at rest and never returned by the API.
Chat model	The model that answers in the chat panel.
Fast model	The model used for lighter supporting tasks.
Reasoning	Whether the chat model reasons before answering, and at which level.

Test Connection confirms the configuration works. It checks that the credential reaches the provider, then checks the entered chat and fast models against the saved connector settings.

Configuring the provider requires the ai:admin entitlement; using the Assistant requires ai:access.

Configuring the AI assistant's model, and the assistant analyzing data in the Studio

Tool Approval

Every Assistant tool has a risk level. The Studio asks for approval before destructive calls; read-only and mutating calls run directly.

Risk	Tools
Read-only	`list_items`, `read_schema`, `read_data_sources`
Mutating	`create_items`, `update_item`
Destructive	`delete_item`, `mutate_schema`

The Assistant runs read-only and mutating tools directly, and stops for a destructive one. When it stops, it names the tool and the change it intends to make, and waits — approve it to run, or deny it to return control without touching your data.

Note

Approval is a second gate, not the only one. A denied permission still blocks a tool the acting user cannot perform, whether or not it is approved.

The MCP Server

The MCP server exposes the same capabilities to external AI agents and tools over the Model Context Protocol. Each workspace has its own endpoint, and agents authenticate with an API key — so an agent acts with exactly that key's permissions. Using it requires the ai:mcp entitlement.

See Configure MCP to connect an agent.

AI