openclaude

Abstracts multiple LLM providers (Claude, OpenAI, local models via Ollama) behind a single agent interface, routing requests based on model availability and configuration. Uses a provider-agnostic message protocol that translates between different API schemas (Anthropic's messages API, OpenAI's chat completions, local inference formats) at runtime, enabling seamless model switching without code changes.

Exposes agent capabilities through a command-line interface that reads task definitions from files, executes agents with file I/O capabilities, and writes results back to the file system. The CLI layer implements a file-watching pattern for continuous agent execution and integrates with shell environments, allowing agents to be triggered from scripts, cron jobs, or CI/CD pipelines without requiring programmatic API calls.

Enables agents to invoke external tools and APIs by registering function schemas that are passed to the LLM, which then decides when and how to call them. Uses a schema-based function registry where developers define tool signatures (parameters, return types, descriptions) once, and the system automatically translates between the agent's tool-call decisions and actual function invocations, handling parameter validation and error propagation.

Implements a planning-reasoning loop where agents break down complex tasks into subtasks, execute them sequentially or in parallel, and adapt based on intermediate results. Uses a state machine pattern where agent state transitions between planning, execution, and reflection phases, with each phase producing artifacts (task lists, execution results, error analyses) that inform subsequent decisions.

Integrates with Ollama to run open-source language models (Llama, Mistral, etc.) locally without cloud API calls. Implements a provider adapter that translates agent requests into Ollama's REST API format, handles model loading/unloading, and manages local inference with configurable parameters (temperature, context window, quantization levels).

Agents can analyze source code by reading files, understanding syntax and structure, and generating code modifications or new implementations. Uses language-specific parsing (likely AST-based for JavaScript/TypeScript) to understand code structure, enabling agents to make targeted edits rather than naive text replacements, and to reason about code semantics (variable scope, function dependencies, type information).

Maintains agent state across multiple invocations, including conversation history, task progress, and learned context. Implements a state persistence layer that serializes agent state (current task, completed steps, tool results) to disk or external storage, enabling agents to resume interrupted tasks and maintain long-term memory of previous interactions.

Implements comprehensive error handling throughout the agent lifecycle, including LLM API failures, tool execution errors, and invalid agent decisions. Uses a fallback strategy pattern where agents can retry failed operations, switch to alternative tools/providers, or escalate to human intervention when recovery is not possible.