Skip to content

Product principles

aire aims to be the agent-first path from idea → evaluated system → deployable API, without forcing a cloud vendor.

Principles

  1. One interface — Prefer AI.* and public exports from aire over deep imports.
  2. Works offline — Core demos, tests, and aire doctor succeed with mock:echo / local:hashing.
  3. Discoverable — Capabilities and honesty notes live in .describe() and docs, not marketing copy alone.
  4. Budgeted & auditable — Agents have step/token/cost limits; steps and sessions are inspectable.
  5. Safety by default — Tool side-effects are classified; policies can deny/require approval; sandbox file tools.
  6. Lazy extras — Optional stacks (serve, torch, neo4j, …) stay out of the critical import path.

Non-goals

  • Not a hosted platform — aire is a library; hosting is your process / containers.
  • Not a full MCP host — see MCP; explicit stdio/HTTP subset.
  • Not pretrained foundation weights — AI.training.foundation(...) builds toy architectures.
  • Not a replacement for specialized ML frameworks — aire.ml is complementary, not PyTorch/sklearn.
  • Not silent magic — misnamed or stub metrics/backends are documented in Honesty.

Maturity

v0.4.0 is Alpha. Core RAG + agents + gateway + eval work offline; several platform extras remain experimental. Read GAPS.md for documented subsets.