Open Claude Code

What does Open Claude Code do?

Open Claude Code describes itself as "An open-source, model-agnostic AI coding agent that lives in your terminal." It runs as a Python CLI, supports multiple providers so a user is not locked to one, and ships 25 built-in tools for file work, code search, shell commands, web search, sandboxed Python, sub-agents and MCP integration.

Three modes shape the loop. Ask is a plain Q&A: the model answers, nothing runs on the machine. Plan reads the repository and writes a plan the user approves before anything executes. Agent runs autonomously under the project's rules, with context-aware compaction so a long session fits the window. Session transcripts are kept locally as JSONL for a durable ledger.

Key facts

• Licence: MIT.
• Platform: anywhere Python 3.12 or later runs. Linux, macOS and Windows.
• Install: pip install open-claude-code from PyPI. Docker is supported for sandboxed runs.
• Providers: Claude (Anthropic), GPT (OpenAI), Gemini (Google), Groq, OpenRouter, Ollama, and any OpenAI-compatible endpoint.
• Tools: 25 built-in tools, listed by category in the README (file ops, Git inspection, web search, sandboxed Python, sub-agents, MCP).
• Modes: Ask, Plan, Agent.
• Skills: YAML with Markdown, readable and editable by the user.
• Model Context Protocol: server integration supported.
• Session ledger: JSONL, kept locally.
• Context handling: compaction powered by a summarising LLM call.
• Cost tracking: per-session tokens and dollars, with budget limits.
• Tests: more than 200 tests covering the agent loop, config, tools and providers, per the README.
• External dependency note: ripgrep (rg) is used for optimised code search; the agent falls back to pure Python without it.

What does it replace, and where does it fall short?

Open Claude Code replaces Claude Code for developers who want a provider-agnostic Python CLI and are happy to pay the model providers directly.

Where it falls short: it is a terminal tool, with no IDE UI of its own. Claude Code itself ships with features that this does not try to match, such as Anthropic's own Managed Agents and plug-ins to Anthropic's hosted products. The project needs Python 3.12 or later, which is newer than some long-term distributions ship by default, and API keys for the chosen provider or a local Ollama install. Cost tracking covers the agent's own budget, not the provider's billing dashboard.

How does Open Claude Code run?

pip install open-claude-code installs it from PyPI. On first run, the agent asks which provider to use and reads the matching API key from the environment (ANTHROPIC_API_KEY, OPENAI_API_KEY, GOOGLE_API_KEY, GROQ_API_KEY, OPENROUTER_API_KEY or a local Ollama endpoint). The README adds a Docker image as an alternative, for a sandboxed run.

Launch the CLI in a project folder and pick a mode. Ask writes a reply. Plan reads the repository and proposes steps, which the user approves one by one. Agent runs under the project's rules until it finishes or hits a budget. Session transcripts are written to a JSONL file in the project.

Who is Open Claude Code for?

A developer who wants a Python coding agent that is not tied to one model provider, who runs from the terminal, and who wants cost tracking and a readable session ledger. A user who wants the native product should install Claude Code itself. A user who wants a smaller Rust binary should look at claude-agent-rs on Linux or macOS.

What limits does the README state?

• Python 3.12 or later is required.
• API keys for the chosen provider, or a local Ollama install.
• Ripgrep (rg) is used for optimised code search; the agent falls back to pure Python without it.
• The 25 built-in tools are the full tool set; others come through MCP.
• Cost tracking is session-level; provider dashboards remain the source of truth for billing.

Questions people ask

Is Open Claude Code open source?

Yes. The LICENSE file at the repository root is MIT, an OSI-approved licence. The code is pure Python and the README names "200+ tests" covering the agent loop, config, tools and providers. The project is distributed through PyPI as open-claude-code.

Does Open Claude Code work offline?

Partly. The agent itself runs on the machine, and the local Ollama provider path keeps the model on the machine too. The hosted providers (Anthropic, OpenAI, Google, Groq, OpenRouter) call their servers for every reply. Web search tools call the web. A reader who needs a full offline run should pair Open Claude Code with an Ollama model.

Can Open Claude Code be used commercially?

Yes for the code, which is MIT. The providers it talks to have their own terms (Anthropic's for Claude, OpenAI's for GPT, Google's for Gemini, Groq's, and so on). Ollama is MIT. A commercial user reads the provider's terms for the model they pick; Open Claude Code itself adds no restriction.

How does Open Claude Code compare with Claude Code?

Open Claude Code is a Python CLI that works with many providers and is MIT-licensed. Claude Code is Anthropic's own product, tied to Anthropic's models, with its own Max subscription and Managed Agents. Open Claude Code is for readers who want provider-agnostic tooling and a readable session ledger; Claude Code is for readers who want Anthropic's fuller feature set. The open-source alternatives to Claude Code page sets it beside OpenCode, Crush, Goose and other provider-agnostic agents.

What does the Plan mode do?

The README's answer is that Plan mode reads the repository, writes a step-by-step plan, and waits for the user to approve each step before anything executes. This is a middle ground between Ask (read-only chat) and Agent (autonomous execution). It is useful on a change where a wrong action is expensive to undo, because the user sees every step on paper before any tool fires.

Sources

• Open Claude Code README and LICENSE: github.com/NastyRunner13/Open-Claude-Code, read 2026-11-09.