Project Philosophies

Core Architecture

• Unified provider interface (adapter pattern). Keep a layered architecture (Specifications → Utilities → Providers → Core) that enables a single, consistent API across many AI providers.

  • This is our central architecture backbone and the heart of what the AI SDK is.
  • It also enables community providers to be developed independently in 3rd party packages.

• Keep the building blocks separated. Building blocks beyond the provider abstraction layer must be cleanly architected and not entangled with it.

  • Critical for tree shaking and agentic development.
  • Enforcing architectural boundaries reduces complexity and the potential for side effects.

• Lean, focused mission. Keep the AI SDK centered on its core mission: the provider abstraction layer, plus directly associated building blocks on top (e.g. the UI chatbot protocol).

  • Be conservative about adding entirely new building blocks. Any such feature needs to be carefully evaluated with the team.
  • The better solution often is to create a separate project built on top of the AI SDK.

API Design

• Stability and backward compatibility first. Changes must remain backward compatible — never change the signature of existing public functions. The only exception is a new AI SDK major release.

  • Even in a major version, breaking changes should have a good justification.
  • If keeping a public API unchanged would result in inferior or painful DX, making the breaking change is absolutely right — it just must happen as part of a new major release.

• Be extremely cautious with @ai-sdk/provider. This package contains the spec. Treat any spec changes as potentially breaking.

  • Ideally, @ai-sdk/provider changes are only made in alignment with a new AI SDK major release.

• Conservative API surface. Keep provider option schemas as restrictive as possible to preserve flexibility for future changes.

  • Keep response schemas minimal (no unused properties).
  • Keep schemas flexible enough to handle provider API changes without unnecessary breakages.
  • Use minimal package exports, especially from the @ai-sdk/provider package, which is responsible for the spec. Usage of the TypeScript primitives Params and ReturnType is encouraged in consuming code over having direct exports of the underlying types.

• Beware premature abstraction. Provider APIs evolve quickly. Avoid adding generic parameters or abstractions that translate differently across providers.

  • Follow the rule of 3: wait until at least 3 providers have implemented the same concept before generalizing, to ensure the abstraction is solid.
  • When unsure or provider-specific, prefer providerOptions.
  • There can be significant pressure to abstract based on one provider. Resist it.

• Use Experimental_ prefixes to explore new features outside of major releases. When a new feature needs to be explored outside of a major release cycle, use code structures explicitly marked as experimental (e.g. Experimental_* prefix for types, experimental_* prefix for functions). This allows iteration without committing to a stable API contract.

  • It is acceptable for @ai-sdk/provider to export Experimental_* types for this purpose. These types may have breaking changes outside of major releases.
  • Non-experimental types must NEVER include references to experimental types (e.g., do not add a reference to something like Experimental_VideoModelV4 to ProviderV4).
  • Experimental features must remain fully isolated until they are promoted to stable.
  • Adding a new experimental feature requires broad consensus between the maintainers. Use it with caution. Do not use experimental code as a way out when you're unsure about stability.

• Clear, accurate naming. When in doubt, prefer longer, more explicit names that are unambiguous and correct (e.g. .languageModel(id) over .chat(id)).

  • Optimize for clarity for both developers and coding agents, not brevity.

Developer & Agent Experience

• Build with developers and agents in mind. Consistent APIs, development patterns, and naming conventions are key.

  • Monitor common agent hallucinations encountered when using agents to write AI SDK code.
  • Agent hallucinations can be worth considering as a suggestion to make the API work the way the agent expected it in the first place.

• DX through consistency. Consistent naming conventions and development patterns improve developer experience.

  • Normalized conventions are extremely critical for coding agents — document them in AGENTS.md.
  • This matters especially across provider implementations that are technically decoupled.