Glossary

Added / Available. The two views on each capability page. Added is what the agent can use in a scope. Available is what connected plugins offer that nobody has added yet.

Agent. A named specialist capability with its own prompt and optionally its own model. Distinct from “the agent”, the assistant that answers in a conversation.

App. A small web app Cengrid hosts behind sign-in, shipped by a plugin or published by the agent.

Base model. The model a conversation runs on. Set per organisation, team or person.

Capability. Anything a turn can reach that an administrator chooses per scope: instruction, skill, agent, MCP server, connection, environment or app.

Connection. One configured account or service identity using a connector, with its credential in the encrypted keychain.

Connector. A description of how Cengrid reaches an external service. Definitions grant no access.

Core. The generic, headless service that runs every turn, enforces identity and policy, and talks to Postgres.

Deployment directory. The committed directory holding cengrid.config.jsonc, sandbox tools and skills, and generated manifests that the CLI deploys.

Egress policy. The rules deciding which hosts calls from the agent and through connections may reach.

Environment. A declaration of the tools installed on a conversation’s computer, using mise.toml and setup.sh.

Instruction. Guidance sent to the model on every turn in the scope it is added to.

Keychain. The encrypted store for credentials. A person’s Keychain page manages their personal connections.

Lock. An organisation setting, Lock for every scope, that pins an item’s on or off state in every team.

Marketplace. A GitHub repository with a marketplace.json listing up to 100 plugins.

MCP server. A remote Model Context Protocol endpoint whose tools Cengrid discovers.

Pi. The agent runtime Cengrid uses to run models from several vendors.

Plugin. A directory in a GitHub repository that ships capabilities, described by a plugin.json.

Portal. The public front door in front of the web UI and admin console, which verifies sign-in.

Posture. The organisation’s security posture: Strict, Auto or Unrestricted.

Principal. A person’s identity, their lower-cased verified work email.

Re-sync. An administrator-triggered fetch that updates a plugin’s definitions from GitHub.

Room. A shared conversation space, room:<platform>:<ref>: a Slack channel, a Google Chat space or a web room.

Sandbox. The durable, isolated computer each scope gets, where the agent runs commands.

Scope. Organisation, team, person or room, the unit that owns state and capabilities.

Skill. A Markdown playbook with supporting files that the agent reads when a task calls for it.

Team. A group of people and rooms sharing capabilities, instructions and a model.