Tools & agents
API (for AI)API
Also known as: application programming interface
A standard way for software to talk to an AI model over the internet, so developers can build the model into their own apps.
An API is the doorway that lets other software use an AI model. Instead of you typing into a chatbot, a program sends its request to the AI's API and gets an answer back automatically. It is how AI gets built into apps, your email assistant, a customer-service bot, a writing tool, all talk to a model through its API behind the scenes.
An API (application programming interface) is a defined way for one piece of software to request services from another. For AI, a model provider exposes an API so developers can send a prompt and receive the model's response programmatically, without hosting the model themselves.
Why it matters
- Integration: it is how AI capabilities get embedded into products and workflows.
- Access model: you call the API and pay per use (usually per token), the model runs on the provider's servers (inference).
- Building block: tool use, agents and RAG systems are all built on top of API calls.
An SDK is a friendlier code library that wraps the raw API.
An AI API exposes model inference (and related services, embeddings, moderation, files) as a network interface, typically HTTP/REST or streaming, with authentication, rate limits and usage-based (per-token) billing. It abstracts hosting and hardware, letting developers consume models without managing GPUs, and standardises request/response schemas (messages, roles, tool/function definitions, structured-output constraints).
Architecturally, APIs are the integration substrate for the entire application layer: orchestration, tool use, RAG and agentic systems are compositions of API calls. Practical concerns include latency and streaming, rate/quota management, retries and idempotency, versioning and model deprecation, data-handling terms (privacy, training-on-inputs), and cost control. Interoperability efforts (compatible schemas, and protocols like MCP for tool/data connections) reduce lock-in. Self-hosting or on-device deployment is the alternative when data residency, cost at scale, or offline operation outweighs API convenience.