OpenTools

Work in progress — this is an early-stage exploration, not a finished spec or production-ready library.

OpenTools is an idea for letting LLMs auto-discover and interact with web app APIs using OpenAPI extensions. Instead of building custom integrations or MCP servers for every app, a client just needs a URL — the rest is automatic.

The Idea

Any web app can expose its API to LLMs by adding x-llm extensions to its OpenAPI spec and serving a /.well-known/llm.json discovery endpoint. Any AI-powered client can then:

1. Fetch the discovery endpoint from a URL
2. Read the OpenAPI spec with x-llm metadata (approval policies, destructive flags, hints)
3. Dynamically generate AI SDK tools — no hardcoded integrations

User: "Create a task called 'Buy groceries' due tomorrow"

→ Chat app already connected to task tracker via URL
→ LLM sees auto-discovered tools from the app's OpenAPI spec
→ LLM calls createTask tool
→ User approves (per-call policy) → task created

Repo Structure

Turborepo monorepo with three library packages and two demo apps:

packages/
  spec/       → @opentools/spec      shared types + constants (zero deps)
  orpc/       → @opentools/orpc      provider adapter for oRPC
  ai-sdk/     → @opentools/ai-sdk    consumer adapter for Vercel AI SDK

apps/
  tasks/      → Task Tracker         demo app exposing an OpenTools-enabled API
  chat/       → Chat App             demo client that connects to any OpenTools app
  web/        → Landing page

Provider side (@opentools/orpc)

Helpers for web apps to annotate their oRPC routes with x-llm metadata and serve the discovery endpoint.

Consumer side (@opentools/ai-sdk)

createToolsFromUrl(url, opts) — takes a URL, fetches the discovery endpoint, parses the OpenAPI spec, and returns AI SDK-compatible tools ready to pass to streamText() or generateText().

Shared spec (@opentools/spec)

Types for the x-llm extensions: approval levels (auto | per-call), destructive flags, cost indicators, discovery document shape.

Demo Apps

Task Tracker (apps/tasks) — simple CRUD app built with oRPC + Next.js. Exposes /.well-known/llm.json and an OpenAPI spec with x-llm annotations on each operation.

Chat App (apps/chat) — AI chat interface using Vercel AI SDK. Connect any OpenTools-compatible app by URL, and the LLM can interact with it through auto-discovered tools with per-action approval.

Status

This is an early exploration of the concept. The core packages (spec, orpc, ai-sdk) have initial implementations. The demo apps are partially built. Auth is API-key only for now (OAuth2 + CIMD planned for later).

Not yet: npm published, production-tested, or spec-finalized.

Tech Stack

• Runtime: Bun
• Build: Turborepo
• Framework: Next.js
• API: oRPC + OpenAPI
• AI: Vercel AI SDK v6
• UI: shadcn/ui + Tailwind CSS
• Code quality: Ultracite (Biome)

Development

bun install
bun dev        # starts all apps (web :3000, tasks :3001, chat :3002)

Docs

See docs/ for more detail:

• demo-goal.md — full demo vision and auth strategy
• package-architecture.md — package design and API surface
• implementation-plan.md — build phases and verification steps