growthxai/output

The open-source TypeScript framework for building AI workflows and agents. Designed for Claude Code describe what you want, Claude builds it, with all the best practices already in place.

What it solves

Output is a TypeScript framework designed to eliminate the fragmentation of the AI stack. Instead of using separate subscriptions for prompt management, tracing, evaluation, and cost tracking, Output integrates these capabilities into a single codebase. It specifically targets the needs of AI coding agents (like Claude Code), ensuring that the entire workflow—including prompts, tests, and traces—is accessible to the agent for easier iteration and development.

How it works

The framework uses a structured approach to build AI workflows:

  • Workflows: The orchestration layer that handles deterministic coordination logic without I/O.
  • Steps: The execution layer where I/O occurs (API calls, LLM requests). Results are cached for replayability.
  • Prompts: Dedicated .prompt files using YAML frontmatter and Liquid templating for version-controlled prompt management.
  • Evaluators: LLM-as-judge systems that provide confidence scores and reasoning to test AI quality.
  • Infrastructure: It leverages Temporal under the hood to provide automatic retries, exponential backoff, and parallel execution.
  • Security: Credentials are encrypted with AES-256-GCM and managed via CLI to keep secrets away from coding agents.

Who it’s for

TypeScript developers and AI engineers who want to avoid vendor lock-in and SaaS fragmentation by keeping their AI infrastructure in their own codebase. It is particularly useful for those using AI coding agents to build and iterate on their AI applications.

Highlights

  • Agent-First Design: Structured for AI coding agents to scaffold, plan, and iterate on workflows.
  • Unified Tooling: Integrated prompts, evals, tracing, and cost tracking in one framework.
  • Multi-Provider Support: A single API for Anthropic, OpenAI, Azure, Vertex AI, and Bedrock.
  • Durable Execution: Built on Temporal for reliable scaling and failure recovery.
  • Local-First Tracing: Traces are stored as JSON in the file system, allowing AI agents to analyze and fix issues directly.
  • Encrypted Credentials: Built-in AES-256-GCM encryption for API keys.