Nerve

Nerve enables agents to be defined as YAML files specifying system prompt, task description, available tools, and LLM parameters, which are then loaded by the runtime system and executed in a loop until task completion. The declarative approach decouples agent logic from execution infrastructure, allowing agents to be version-controlled, audited, and reproduced deterministically without code changes.

Nerve abstracts multiple LLM providers (OpenAI, Anthropic, Ollama, etc.) behind a unified interface, allowing agents to switch providers by changing a single configuration parameter without code changes. The runtime system handles provider-specific API calls, token counting, and response parsing transparently.

Nerve implements an agentic loop where the LLM is repeatedly prompted with the current task state and available tools, generates tool invocations or task completion signals, and the runtime executes tools and updates state. The loop continues until the LLM signals task completion or a maximum iteration limit is reached, with all invocations logged for auditability.

Nerve's tool system provides agents access to three categories of tools: shell commands executed in subprocess, Python functions loaded from modules, and remote tools exposed via MCP protocol. Tools are registered in namespaces with JSON schemas describing inputs/outputs, enabling the LLM to invoke them with proper argument validation and error handling.

Nerve workflows enable sequential chaining of multiple agents where each agent executes in order and passes shared state to the next agent via a state dictionary. The workflow runtime manages state propagation, handles inter-agent dependencies, and provides a single execution context for the entire workflow. Agents can read and modify shared state, enabling data flow and coordination between steps.

Nerve implements both MCP client and server modes, allowing agents to consume remote tools from MCP servers and expose their own tools to other agents via MCP. The MCP integration uses standard MCP protocol for tool discovery, schema negotiation, and remote invocation, enabling tool sharing across agent boundaries without code coupling.

Nerve provides an evaluation system that runs agents against predefined test cases, comparing actual outputs against expected results and collecting performance metrics. The evaluation framework supports multiple test formats, tracks success/failure rates, and enables benchmarking agents across different configurations or LLM providers to measure improvement over time.

Nerve's runtime maintains a state dictionary that persists across agent execution steps and workflow stages, allowing agents to read previous results, accumulate data, and coordinate through shared context. The state system provides isolation between workflow runs while enabling transparent data flow between sequential agents without explicit serialization.

Nerve provides a CLI command that guides users through creating new agents interactively, prompting for system prompt, task description, available tools, and LLM configuration. The wizard generates a valid YAML agent definition file, reducing friction for new users and ensuring agents follow Nerve conventions.

Nerve includes built-in tool namespaces providing common utilities like file operations, HTTP requests, and system commands without requiring custom tool implementation. These namespaces are pre-configured with appropriate schemas and error handling, allowing agents to immediately access common capabilities.

Nerve provides a CLI command to execute agents defined in YAML files, with support for overriding configuration parameters via command-line arguments. The CLI handles agent loading, LLM initialization, tool registration, and execution loop management, providing a simple entry point for running agents without writing Python code.