@tinyweb-agent-framework/llm

LLM core: schemas, OpenAI Chat protocol, route architecture, tool runtime, and provider profiles.

Adapted from opencode's @opencode-ai/llm package, slimmed to OpenAI Chat only.

Surface

import {
  LLM,                  // request/generate/stream/helpers facade
  LLMClient,            // Service + layer + prepare/stream/generate
  Tool,                 // Tool.make typed-tool builder
  Route,                // Route.make / Route.model
  OpenAI,               // canonical OpenAI provider
  OpenAICompatible,     // deepseek/together/groq/cerebras/fireworks/deepinfra/openrouter/xai/ollama/lmstudio
  Auth, AuthOptions,    // bearer / header / config / fromModelApiKey
  ProviderShared,       // joinText, parseToolInput, validateWith, …
} from "@tinyweb-agent-framework/llm"

Models

import { OpenAI, OpenAICompatible } from "@tinyweb-agent-framework/llm"

// OpenAI
const m1 = OpenAI.chat("gpt-4o-mini", { apiKey: process.env.OPENAI_API_KEY })

// Any OpenAI-compatible profile (DeepSeek, Together, Groq, Ollama, …)
const m2 = OpenAICompatible.deepseek("deepseek-chat", { apiKey: process.env.DEEPSEEK_API_KEY })
const m3 = OpenAICompatible.ollama("llama3", { apiKey: "ollama" })

Tools

import { Tool, ToolFailure } from "@tinyweb-agent-framework/llm"
import { Effect, Schema } from "effect"

const search = Tool.make({
  description: "Search the web",
  parameters: Schema.Struct({ query: Schema.String, limit: Schema.optional(Schema.Number) }),
  success: Schema.Struct({ results: Schema.Array(Schema.String) }),
  execute: ({ query }) =>
    Effect.tryPromise({
      try: () => fetch(`https://example.com/api?q=${query}`).then((r) => r.json()),
      catch: (e) => new ToolFailure({ message: String(e) }),
    }),
})

Architecture

Four-axis decomposition: Protocol (body & stream schema), Endpoint, Auth, Framing — composed via Route.make({ id, protocol, transport, defaults }). See src/route/ and src/protocols/openai-chat.ts.

License

MIT