AgentMark Prompt Core

The core prompt engine for AgentMark. Parses .prompt.mdx files and formats them into a neutral { messages, ...config } shape you hand to whatever LLM SDK you already use — the Vercel AI SDK, the raw OpenAI or Anthropic client, Pydantic AI, or your own bespoke client.

Installation

npm install @agentmark-ai/prompt-core

Quick Start

import { createAgentMark } from "@agentmark-ai/prompt-core";
import { FileLoader } from "@agentmark-ai/prompt-core/loader-file";

// Loads prompts pre-built by `agentmark build`
const agentmark = createAgentMark({
  loader: new FileLoader("./dist/agentmark"),
});

const prompt = await agentmark.loadTextPrompt("customer-support.prompt.mdx");
const { messages, ...config } = await prompt.format({
  props: { customer_question: "How long does shipping take?" },
});

// Hand `messages` + `config` to your SDK of choice

See the bring-your-own-SDK guide for the full integration path.

API

createAgentMark(options)

Create an AgentMark instance.

const agentmark = createAgentMark({
  loader,          // Optional: where prompts load from (file, API, or custom)
  evals,           // Optional: eval functions keyed by name, for experiments
  builtInModels,   // Optional: allowed model names (validates frontmatter model_name)
  templateEngine,  // Optional: custom template engine (defaults to TemplateDX)
});

Loading prompts

One loader method per generation type. Each returns a typed prompt object with a format() method:

• agentmark.loadTextPrompt(path) — text generation (text_config)
• agentmark.loadObjectPrompt(path) — structured output (object_config)
• agentmark.loadImagePrompt(path) — image generation (image_config)
• agentmark.loadSpeechPrompt(path) — speech generation (speech_config)

Loaders

• FileLoader (@agentmark-ai/prompt-core/loader-file) — load prompts pre-built by agentmark build. For self-hosted deployments with no runtime API calls.
• ApiLoader (@agentmark-ai/prompt-core/loader-api) — load prompts from AgentMark Cloud or a local agentmark dev server.

See Loaders.

The standalone @agentmark-ai/loader-api and @agentmark-ai/loader-file packages are re-export shims of these subpaths — prefer the subpaths in new code.

Type safety

Generate types from your prompts with agentmark generate-types, then parameterize the instance for end-to-end type-safe props and outputs:

import type AgentmarkTypes from "./agentmark.types";

const agentmark = createAgentMark<AgentmarkTypes>({ loader });

See Type safety.

Documentation

Full documentation at docs.agentmark.co.

License

AGPL-3.0-or-later