Glossary
Common Brain terms.
| Term | Meaning |
|---|---|
| Agent mode | Chat mode with MCP tools and sandbox coding |
| Ask mode | Plain chat without tools |
| Attached repo | GitHub repo cloned into the sandbox for coding |
| BYOA | Bring your own app credentials (workspace, host, or env) — also used for custom model endpoints |
| BYOK | Bring your own model API key |
| Command Code | Default OpenAI-compatible model provider (curated catalog) |
| Connection | An MCP app Brain can call |
| DCR | Dynamic client registration (no static client secret) |
| dFlow Enterprise | Dokku-based host commonly used to deploy Brain |
| eve | Agent runtime Brain proxies via /eve/v1/* |
| Foundry | Azure AI Foundry — private model plane; register as a custom model |
| HITL | Human-in-the-loop approval before a tool runs. Frequency follows instance agent safety posture (default Auto: writes wait, many reads skip). |
| Instance admin | Host-level admin (/settings/instance) |
| MCP | Model Context Protocol — live tool calls, not a document index |
| microsandbox | Sandbox backend for attached-repo coding (not Vercel Sandbox) |
| Playbook | Saved prompt shortcut |
| RAG | Retrieval-augmented generation over a copied index — not how Brain reads work systems |
| Result screening | After a tool runs, Auto/Strict replace injection- or secret-like output with a stub before the model sees it. Not an approve prompt. Dangerous live chats skip it. |
| Sandbox | Isolated /workspace for file edits during a coding turn |
| SCIM | Directory provisioning API (/api/auth/scim/v2, Users only) |
| Workspace | Tenancy boundary for chats, tools, and members |