What is Flint?

Flint is a token-efficient agentic TypeScript runtime. It gives you a minimal, composable set of typed building blocks for AI agents — and then stays out of the way.

The core idea

Most AI agent frameworks abstract away the LLM interaction behind chains, classes, and hidden state. Flint inverts this: you get six plain functions and compose them yourself using ordinary TypeScript. The framework doesn't run your agent — JavaScript does.

import { call, tool, agent } from 'flint';
import { budget } from 'flint/budget';
import { anthropicAdapter } from '@flint/adapter-anthropic';

What Flint is not

• Not a RAG framework — RAG utilities are included but minimal; bring your own vector database.
• Not an orchestration platform — no server, no deployment, no hosted runtime.
• Not opinionated about prompt engineering — Flint doesn't template your prompts.

Packages

Package	Role
flint	Core: primitives, agent loop, budget, compress, memory, RAG, recipes, safety
@flint/adapter-anthropic	Anthropic Messages API — prompt-cache aware
@flint/adapter-openai-compat	Any OpenAI-compatible endpoint
@flint/graph	State-machine agent workflows

Design principles

One runtime dependency. The flint core depends only on @standard-schema/spec. Adapters have zero runtime dependencies.

Standard Schema for tool inputs. Use Zod, Valibot, ArkType, or any Standard Schema-compatible library for tool input validation. Flint doesn't bundle a schema library.

Results, not exceptions. All public async functions return Result<T> — a discriminated union { ok: true; value: T } | { ok: false; error: Error }. No try/catch at the call site.

Web API primitives only. Requires Node 20+ but uses only fetch, ReadableStream, and TextDecoder — works in edge runtimes.