llm-schemas

Shared Zod schemas for validating LLM API request bodies. Covers OpenAI Chat Completions, OpenAI Responses, and Anthropic Messages. Used by both llm-mock-server and copilot-sdk-proxy.

Table of contents

• Install
• Usage
  • OpenAI Chat Completions
  • OpenAI Responses
  • Anthropic Messages
  • Top-level import
• What's in each schema
  • Request schemas
  • Building blocks
• Import paths
• Development
• Licence

Install

npm install llm-schemas

Requires Node.js 22+ and Zod 4.

Usage

Import from a sub-path to get schemas for a specific API, or import from the top level to get everything namespaced.

OpenAI Chat Completions

import {
  OpenAIRequestSchema,
  MessageSchema,
  ToolSchema,
} from "llm-schemas/openai/chat-completions";

const body = OpenAIRequestSchema.parse(req.body);
console.log(body.model, body.messages);

OpenAI Responses

import {
  ResponsesRequestSchema,
  FunctionToolSchema,
  InputMessageSchema,
} from "llm-schemas/openai/responses";

const body = ResponsesRequestSchema.parse(req.body);
console.log(body.model, body.input);

Anthropic Messages

import {
  AnthropicRequestSchema,
  MessageSchema,
  ToolDefinitionSchema,
} from "llm-schemas/anthropic";

const body = AnthropicRequestSchema.parse(req.body);
console.log(body.model, body.max_tokens, body.messages);

Top-level import

If you need all three in one place, the root export namespaces them so there are no naming collisions.

import { openai, anthropic } from "llm-schemas";

openai.chatCompletions.OpenAIRequestSchema.parse(body);
openai.responses.ResponsesRequestSchema.parse(body);
anthropic.AnthropicRequestSchema.parse(body);

What's in each schema

Request schemas

All three request schemas use z.looseObject() at the top level, so unknown fields pass through without failing validation. Every field from the official API specs is included. Fields that don't need parsing are typed as z.unknown().optional().

Specs the schemas were built from:

• OpenAI Chat Completions
• OpenAI Responses
• Anthropic Messages

| Schema | Required fields | Total fields | | ------ | --------------- | ------------ | | OpenAIRequestSchema | model, messages | 35 | | AnthropicRequestSchema | model, max_tokens, messages | 18 | | ResponsesRequestSchema | (none) | 29 |

The Responses API spec makes both model and input optional. If your app needs them to be required, add a .refine() after parsing.

Building blocks

OpenAI Chat Completions

• ContentPartSchema - a content part like { type: "text", text: "..." } (looseObject)
• MessageSchema - a chat message with role, content, tool_calls, etc.
• ToolSchema - a tool definition with function.name, function.parameters, function.strict

Anthropic Messages

• TextBlockSchema, ToolUseBlockSchema, ToolResultBlockSchema - typed content blocks
• LooseContentBlockSchema - union of known blocks + a fallback for unknown types
• MessageSchema - a message with role and string or block array content
• ToolDefinitionSchema - a tool with name, description, input_schema

OpenAI Responses

• InputMessageSchema - a conversation message in the input array
• FunctionCallInputSchema, FunctionCallOutputSchema - function call round-trip items
• FunctionToolSchema - a function tool (with strict field)
• RawToolSchema - accepts any tool shape as Record<string, unknown>

Every schema also exports its inferred TypeScript type (e.g. OpenAIRequest, AnthropicRequest, Message, FunctionTool).

Import paths

| Path | What you get | | ---- | ------------ | | llm-schemas | Everything, namespaced as openai.chatCompletions, openai.responses, anthropic | | llm-schemas/openai | Both OpenAI APIs, namespaced as chatCompletions and responses | | llm-schemas/openai/chat-completions | Chat completions schemas only | | llm-schemas/openai/responses | Responses schemas only | | llm-schemas/anthropic | Anthropic schemas only |

Development

npm run build # Compile TypeScript
npm test # Run tests
npm run lint # Lint with oxlint
npm run check # All three: typecheck + lint + test

Licence

MIT