LLM agents as YAML files you can actually test in CI

What it does

Nerve is an Agent Development Kit that lets you define LLM-based agents in a single YAML file—system prompt, task, tools, and variables included—and run them from the command line. It supports shell commands, Python functions, and remote tools as fully typed capabilities, and it can chain agents into simple linear workflows with shared context. Under the hood it uses LiteLLM, so switching between OpenAI, Anthropic, Ollama, or other providers is a one-line change.

The interesting bit

The project claims to be the first framework that lets you define MCP servers in YAML, acting as both client and server so agents can call each other in teams. That turns a configuration file into an interoperable service endpoint, which is either a neat shortcut or a recipe for indentation-related headaches depending on how you feel about YAML.

Key highlights

• Declarative agent definitions in one YAML file—prompts, tasks, tools, and variables.
• Native MCP client and server support, including YAML-defined MCP servers for agent-to-agent orchestration.
• Built-in evaluation mode with YAML, Parquet, or folder-based test cases for reproducible benchmarking.
• Linear workflow pipelines that share context across multi-step automations.
• LLM-agnostic via LiteLLM, supporting OpenAI, Anthropic, Ollama, and dozens more.

Caveats

• Workflows are explicitly simple and linear, so don’t expect complex DAG-style orchestration or conditional branching out of the box.
• The “first framework” boast for YAML-defined MCP servers is specific enough that you may want to verify it against your own definition of “framework.”

Verdict

Grab it if you want version-controlled, testable LLM agents without writing a Python orchestration layer. Skip it if you need visual drag-and-drop flows or heavy asynchronous DAG execution.