A coding agent CLI that refuses to pick sides

What it does OpenClaude is a terminal-based coding agent that talks to nearly any model provider you already use—OpenAI-compatible APIs, Gemini, GitHub Models, Codex, Ollama, and several others—through one consistent CLI. It handles streaming prompts, tool calling, file edits, bash commands, and MCP servers without caring who hosts the weights.

The interesting bit The project began as a Claude-oriented tool but grew into a universal adapter: it can route different agents to different models (e.g., a cheap flash model for exploration, GPT-4o for planning) and even exposes its engine as a headless gRPC service so other apps or CI pipelines can borrow its tool loop. That turns a chat CLI into a reusable backend component.

Key highlights

• Provider profiles saved inside the app via /provider, so switching from Ollama to Gemini is a command away
• Bundled VS Code extension for launch integration and theme support
• DuckDuckGo web search works out of the box for non-Anthropic models; Anthropic backends keep native search behavior
• Agent routing lets you split work across models by task type for cost or capability optimization
• Headless gRPC mode with bidirectional streaming for embedding the agent into external UIs or automation

Caveats

• Anthropic-specific features don’t necessarily translate to other providers, and smaller local models can choke on long multi-step tool flows
• API keys in ~/.openclaude.json are stored in plaintext, so keep the file private
• DuckDuckGo web search is essentially scraping and may be rate-limited; Firecrawl is the more reliable paid alternative

Verdict Worth a look if you want a single coding agent interface across multiple LLM providers without managing a tangle of separate CLIs. Skip it if you’re all-in on one proprietary ecosystem and don’t need the abstraction.