Installation

Glove is a set of small packages rather than one framework install. You add the runtime, the binding for your UI, and whichever capability packages the app actually needs. Nothing else is pulled in.

Requirements

• Node.js 18+ (20+ recommended — several packages use the modern fetch and web streams).
• TypeScript 5+, ESM. Every package ships "type": "module" with bundled types.
• zod — tool input schemas are Zod schemas. Install it alongside the runtime.
• An API key for at least one model provider, or a local runtime (Ollama / LM Studio) that needs none.

Pick your shape

Four common shapes. Pick the row that matches what you are building — the rest of the docs assume one of them.

# Full-stack (Next.js)
pnpm add glove-react glove-next zod

# Server-only (Node / CLI / worker)
pnpm add glove-core zod

# npm and yarn work the same way
npm install glove-react glove-next zod

glove-core is a dependency of glove-react, so a full-stack install already has the runtime. Install it explicitly when you import from it directly on the server.

Capability packages

Everything else is opt-in and installed the same way. Each one has its own page — this is the map.

Model providers

One adapter factory covers every provider. Pass the provider name and a model; the key is read from the environment.

import { createAdapter } from "glove-core/models/providers";

export const model = createAdapter({
  provider: "anthropic",              // see the table below
  model: "claude-sonnet-4-20250514",
  stream: true,
});

ANTHROPIC_API_KEY=sk-ant-...
# or
OPENAI_API_KEY=sk-...

Local models

Ollama and LM Studio need no key. Point the adapter at the local server and name the model you pulled:

const model = createAdapter({
  provider: "ollama",
  model: "llama3",
  baseURL: "http://localhost:11434/v1", // the default
});

Reasoning and prompt caching

Two switches worth knowing at install time, because they change cost and latency more than anything else you configure:

// Capture provider reasoning traces (DeepSeek, Qwen3-Thinking, GLM, Kimi,
// MiniMax, o-series…). `true` for defaults, or an object to tune depth.
createAdapter({ provider: "openai", reasoning: { effort: "high" } });

// Provider prompt caching — breakpoints on the stable prefix and latest turn.
createAdapter({ provider: "anthropic", cache: { ttl: "1h" } });

// The same switches exist on the Next.js handler.
createChatHandler({ provider: "anthropic", cache: true });

Cache usage is reported on every response (cache_creation_input_tokens / cache_read_input_tokens) and threaded through to useGlove().stats, so you can bill on it. Full detail in the Core API reference.

Building from source

The repo is a pnpm workspace. Clone it if you want to run the examples or work on a package:

git clone https://github.com/porkytheblack/glove.git
cd glove
pnpm install
pnpm build       # build every package
pnpm typecheck

Next: build a working agent in 15 minutes, or skim All Packages to see what is available before you commit to a shape.