Documentation
View .mdOptimized for LLMs — paste directly into ChatGPT, Claude, or Cursor.
Package architecture
Agntz is a monorepo, but the packages have distinct jobs. Use the smallest package that matches your use case.
Public package map
| Package | Use when you need |
|---|---|
@agntz/sdk | Embedded TypeScript execution, local client resources, CLI, local SQLite helper |
@agntz/client | Hosted or self-hosted HTTP client for normalized runs, streaming, rich content, artifacts, retention, sessions, traces, datasets, and evals |
@agntz/core | Low-level runner, tool definitions, model providers, telemetry, MCP helpers, and the manifest parser plus published JSON Schema |
@agntz/contracts | Shared store/resource ports, entity types, namespace grants, agent refs, errors, and leaf utilities |
@agntz/db | Shared SQLite/Postgres connection and migration plumbing |
@agntz/stores | Store contracts plus in-memory, SQLite, and Postgres implementations |
@agntz/memrez | Durable memory resource provider, memory stores, reasoners, curation |
agntz | Python hosted client, embedded runtime, stores, and memrez resources |
The old standalone manifest package has been merged into @agntz/core/manifest.
Boundary rules
- Runtime execution belongs in
@agntz/core. - Shared data contracts and pure helpers belong in
@agntz/contracts. - Database connection/migration helpers belong in
@agntz/db. - Hosted identity and tenant policy belong in app/worker code, not in resource libraries.
- Resource packages such as memrez depend on contracts and accept host-provided model providers instead of constructing the runtime.
- Store adapters implement contracts without depending on core runtime internals.
Which package should I import?
For most applications:
Use @agntz/core directly when you are building a custom runtime host or need the low-level runner. Use @agntz/stores when you need persistent runtime or hosted control-plane storage.
For a backend that would otherwise import a provider SDK, start with @agntz/client and the provider-replacement guide. The Python agntz package exposes the corresponding sync and async hosted clients.