The Claude Code escape hatch that runs on Ollama and GPT-4o alike

What it does

OpenClaude Improved is a community-built CLI coding agent derived from Claude Code. It gives you the same chat-driven, tool-looping workflow—bash commands, file edits, repo mapping, MCP tools—but lets you point it at OpenAI, Gemini, Ollama, OpenRouter, GitHub Models, or a long tail of niche providers instead of being tied to Anthropic’s API. The project bundles a VS Code extension, background session support, and a gRPC server for headless CI use.

The interesting bit

The real work is in the provider abstraction: it normalizes tool loops, streaming, vision, and context windows across twenty-plus backends, including oddballs like Xiaomi MiMo and Z.AI GLM. It even auto-negotiates Ollama’s native chat API to preserve full session history rather than letting the OpenAI compatibility shim silently truncate tokens.

Key highlights

• Provider-agnostic by design: OpenAI-compatible, Gemini, Ollama, Bedrock, Vertex, Codex, and a dozen smaller gateways all supported via the same /provider profile system.
• Repo map uses PageRank to rank your codebase’s structure and auto-injects it behind a flag when context is needed.
• Background sessions run as plain local child processes—no daemon, no network service—with logs and metadata stored in ~/.openclaude/bg-sessions/.
• Includes a VS Code extension and a bidirectional gRPC streaming server for custom UIs or automated pipelines.
• A truecolor pixel-art “buddy” (/buddy) fires signature moves on every Enter, because terminal morale matters.

Caveats

• Anthropic-only features don’t exist on every backend, and tool quality degrades on small local models during long multi-step loops.
• The codebase is derived from Claude Code; Anthropic retains ownership of the original code, and this is an independent (and unaffiliated) community project.
• Some providers cap output below CLI defaults, and not all support vision or extended context windows.

Verdict

Worth a look if you want Claude Code’s agentic workflow but need to use your own API keys, local Ollama instances, or non-Anthropic enterprise endpoints. Skip it if you want an officially supported, polished product with a single vendor to blame when things break.