Home

Glossary

Quick definitions for terms used throughout the docs. Each entry links to the canonical page where applicable.

Admin — The highest assignable role. Full system access including app settings and Admin role assignment. See Roles and scopes.

Alias — A named handle for an inference target. Bodhi has three alias kinds — User aliases (a GGUF file plus inference parameters), Model aliases (auto-generated from a model file), and API model aliases (a remote provider plus credentials). See Models, aliases, files.

API model — An alias backed by a remote provider (OpenAI, Anthropic, Gemini, Groq, OpenRouter, or any OpenAI-compatible base URL) rather than a local GGUF file. Uses the same /v1/chat/completions interface. See API models.

API token — A long-lived credential of the form bodhiapp_<random>.<client_id> used to authenticate programmatic requests. Carries a scope (User or PowerUser). See API tokens.

App access request — A flow where a third-party app requests access to specific MCP instances and APIs on a user's Bodhi. The user reviews and approves resource-by-resource. See App access management.

Bodhi-native envelope — Bodhi's own error JSON shape, used by /bodhi/v1/*. A strict superset of the OpenAI error envelope. See Error format.

BODHI_HOME — Root directory for all Bodhi state — settings, databases, logs, and secrets. Defaults to ~/.bodhi on desktop and /data/bodhi in the Docker images. See Environment variables.

DCR (Dynamic Client Registration) — RFC 7591 flow that lets Bodhi register itself as an OAuth client with an MCP server at connect time, with no manual setup. Requires the server to expose RFC 8414 metadata. See MCP auth methods.

Embedding — A fixed-length vector representing the semantic content of a text. Bodhi exposes embeddings via /v1/embeddings for both local models and configured API models. See OpenAI embeddings.

External app — An app that has been granted resource consent through the app access request flow. Receives an OAuth token tagged with the specific MCPs and API models it can use. See App access management.

GGUF — The on-disk file format used by llama.cpp for quantized model weights. Bodhi downloads GGUF files from HuggingFace and loads them into the embedded llama.cpp server. See Model files.

Guest — Authenticated user without an assigned role. Cannot reach feature pages until an Admin or Manager approves their access request. See User access requests.

HuggingFace cache — The directory under $HF_HOME/hub where downloaded GGUF files live. Shared across local tools that respect HF_HOME. See Model downloads.

llama.cpp — The embedded C++ inference engine that runs local GGUF models. Bodhi spawns a llama.cpp server subprocess per loaded model. See Inference stack.

Manager — Role between PowerUser and Admin. Can approve user access requests and manage user roles up to and including Manager. Cannot promote/demote Admins or edit app settings. See Roles and scopes.

MCP (Model Context Protocol) — Open protocol for tool servers that LLMs can call mid-conversation. Bodhi connects as a client to MCP servers and exposes their tools to the chat UI and the API. See MCP overview.

MCP instance — A user-scoped configured connection to an MCP server, including auth credentials and a per-instance tool whitelist. See MCP setup.

MCP playground — UI at /ui/mcps/playground/ for invoking individual MCP tools with arbitrary inputs to inspect their responses. See MCP playground.

MCP proxy — Endpoint at /bodhi/v1/apps/mcps/{id}/mcp that lets external apps route their MCP traffic through Bodhi's auth gateway. See MCP proxy.

MCP server (catalog) — An admin-curated, workspace-wide MCP server template that users can spin instances from in one click. Stored in the MCP Store. See Pre-registered MCP servers.

MCP store — Admin-only catalog UI for adding, editing, and removing pre-registered MCP servers. See Pre-registered MCP servers.

Model file — A downloaded GGUF artifact stored in the HuggingFace cache. Identified by (repo, filename). See Model files.

OAuth2 PKCE — Proof Key for Code Exchange flow used for browser logins to Bodhi's identity provider (id.getbodhi.app by default). No password is stored in Bodhi itself. See Auth and roles.

PowerUser — Role above User. Can download models, create aliases, configure API models, mint API tokens, and register external apps. Cannot approve user access requests or manage roles. See Roles and scopes.

Quantization — Lossy compression of model weights from float to int. GGUF files come in many quants (Q4_K_M, Q5_K_S, Q8_0, etc.) trading quality for size and speed. See Inference stack.

Refresh token — OAuth token used to obtain a new access token without re-authenticating. Bodhi handles refresh transparently for sessions and stored MCP credentials.

ResourceRole — Bodhi's term for the assignable role on a session: Anonymous, Guest, User, PowerUser, Manager, Admin. See Roles and scopes.

Resource consent — Per-resource grant given by a user when approving an app access request. The app can only act on resources it was granted (specific MCP instances, specific API models). See App access management.

Role — One of the assignable capability tiers on a Bodhi user. See Auth and roles and Roles and scopes.

Scope — A capability ceiling carried by an API token (User or PowerUser). The token's effective capability is min(user_role, scope). See Roles and scopes.

Session cookie — HttpOnly, Secure, SameSite cookie set after browser OAuth login. Used by the built-in UI and refreshed silently. See Auth and roles.

Setup wizard — One-time first-run flow that resolves BODHI_HOME, picks the inference variant, and seats the first Admin. See Install.

System tray — Tauri desktop's persistent icon for launching the UI, opening the home directory, and quitting. See Desktop deployment.

Tauri — The cross-platform shell that wraps Bodhi's HTTP server and embedded UI into a single native desktop app. See Desktop deployment.

Tool whitelist — Per-MCP-instance allow-list of tool names that can actually be invoked from chat or the proxy. Tools not on the whitelist are hidden. See MCP setup.

User — The lowest assignable role. Can chat, call /v1/chat/completions, list models, and manage their own MCP instances. See Roles and scopes.

User access request — A flow where a newly-logged-in user requests a role from an Admin or Manager. See User access requests.