agent-card-runtime-adapters

Turn a Kinetic Gain Agent Card into the agent config your runtime actually wants — OpenAI Assistants, Anthropic Messages (system-prompt bundle), Vercel AI SDK, or a human-readable markdown briefing. One disclosure, multiple deployments.

Pairs with agent-tool-adapters in the Kinetic Gain lane #4 (agent-runtime adapters): that handles per-tool wiring; this handles the whole agent. Grounded on the agent-cards-spec (agent_card_version 0.1).

Why

A Kinetic Gain Agent Card already declares who the agent is, what it's for, which models and tools it uses, how autonomous it is, and what it refuses — exactly the information you would otherwise scatter across system prompts, dashboard configs, and runtime initializers. This library lifts that single source of truth into the shape each runtime expects, so the disclosure isn't a doc artifact that drifts from production: it is the production config.

Pure transform. No network, no LLM, no execution.

Install

npm install -g agent-card-runtime-adapters   # CLI
npm install agent-card-runtime-adapters      # library

Requires Node ≥ 20.

CLI

agent-card-adapt agent-card.json --target openai
agent-card-adapt agent-card.json --target anthropic --out anthropic.json
agent-card-adapt agent-card.json --target markdown   # human briefing for review/onboarding

agent-card.json is a v0.1 Kinetic Gain Agent Card. Exit codes: 0 ok, 2 usage/IO error.

Library

import { convert } from "agent-card-runtime-adapters";

const openaiCfg = convert(card, "openai");       // OpenAI Assistants create payload
const anthCfg   = convert(card, "anthropic");    // { model, system, metadata }
const vercelCfg = convert(card, "vercel");       // { model, system, tools, metadata }
const briefing  = convert(card, "markdown");     // string

What gets mapped

Card field	Runtime use
agent.name / description	OpenAI Assistants name; opening line of the system prompt
capabilities.models_used	picks the planner role first, else the first entry, for the runtime's model
capabilities.primary_purpose + autonomy_level + memory_persistence	the rendered system prompt (used by anthropic/vercel/markdown; OpenAI instructions)
capabilities.tools[]	function envelopes (openai), keyed record (vercel), markdown bullets
refusal_taxonomy[]	refusal block appended to the system prompt + dedicated section in the markdown briefing
agent_card_version / agent.id / agent.version / autonomy_level	propagated to metadata

Tool schemas are intentionally not invented here — Agent Cards reference tools by name; pair this with agent-tool-adapters for full per-tool wiring from MCP tools/list.

License

AGPL-3.0-or-later — see LICENSE.