Capabilities ​
Core is split into focused package entry points. The root export contains common authoring APIs; advanced and environment-specific APIs remain on subpaths.
| Capability | Entry point | Package responsibility |
|---|---|---|
| Agents | @anvia/core/agent | Bounded model/tool loops, sessions, typed interactions and continuations, observability, dynamic context, agent-as-tool |
| Completions | @anvia/core/completion | Provider-neutral requests, streams, messages, usage, documents, structured parsing, typed model controls |
| Tools | @anvia/core/tool | Zod validation, approvals, structured questions, middleware, dynamic discovery, result normalization |
| Documents | @anvia/core/documents | Shared text-document records and deterministic single/batch text chunking |
| Lifecycle | @anvia/core/agent | Typed run, turn, generation, tool, finish, and error observation |
| Guardrails | @anvia/core/guardrails | Input/final-output allow, block, and rewrite policies |
| Memory | @anvia/core/memory | Conversation store, inspection, save policy, durable compaction contracts |
| Retrieval | @anvia/core/embeddings, @anvia/core/vector-store | Raw-text ingestion, dense/sparse embedding helpers, filters, in-memory search, search-tool contracts |
| Pipelines | @anvia/core/pipeline | Typed composition, parallel branches, batch execution, graphs, run observers |
| Extraction | @anvia/core/extractor | Completion-backed schema extraction with retry support |
| Media | @anvia/core/image-generation, speech-generation, transcription | Provider-neutral helpers and model interfaces |
| MCP registration and skills | @anvia/core/mcp, @anvia/core/skills | Lightweight MCP server/tool contracts and validated local skill loading |
| Observability | @anvia/core/observability | Run, generation, tool, trace, and streaming observer contracts |
| Evaluations | @anvia/core/evals | Cases, metrics, suites, reporters, traces, CLI results, cost/usage totals |
| Browser-safe interactions | @anvia/core/agent/interactions | JSON-safe approval/question contracts, assertions, and parsers without loading the Agent runtime |
Capability boundaries ​
Core does not call a model without an injected model implementation. It does not ship durable database or vector infrastructure, authenticate users, authorize tools, schedule background jobs, or host a browser UI.
Provider support also varies. A completion adapter may not implement embeddings, images, audio, transcription, model listing, provider tools, or every structured-output feature. Use the provider capability matrix before assuming parity.
MCP connections and transports live in the optional Node.js package @anvia/mcp. Client messages, the framed wire protocol, and UI conversion live in @anvia/client. Core's document helpers accept app-supplied text and do not discover, read, or parse files. Local skill loading still belongs on a server with filesystem access; ReadableStream helpers need Web Streams support.
For exact exports, see the API reference. For composition guidance, use architecture and patterns.