Open Claude Code tracks upstream nightly by nightly

What it is

An open source CLI that mirrors the architecture of Anthropic's Claude Code: the agent loop, the tool set, the permission modes, the settings chain and the session model. It is not a copy of the source; it is a rebuild informed by decompiled bundles.

What changed

Nightly 2.0.0 releases are published as upstream moves; the repository records the last known upstream version it tracks, currently 2.1.269 as of this week's commits.

The surface is 25 tools, 40 commands, four MCP transports including streamable HTTP, and six permission modes.

Get started

Node.js. No install required:

npx @ruvnet/open-claude-code "what files are in this directory?"

Expected result: the agent answers using its read tools against your current directory.

Connect additional MCP servers the usual way:

{
  "mcpServers": {
    "<server-name>": {
      "command": "npx",
      "args": ["-y", "<published-mcp-package>"]
    }
  }
}

The block above is the config shape, not a package you can install. Substitute a server that actually exists, for example the AgentDB server registered with claude mcp add agentdb -- npx agentdb@latest mcp start.

Run it first in a scratch directory. It is a coding agent with tools; the permission mode you start in decides what it can touch.

Use it today

Practical case: understanding how a modern coding agent is actually put together. Input is a prompt; the workflow is read the loop while it runs; the output is a mental model you can apply to your own harness.

Acceptance test: ask it to list files, then ask it to write one while in a read only permission mode. The second request should be blocked.

Push it further

Experimental commentary. An open agent loop is a place to test ideas an upstream vendor will not ship — alternative routers, different memory, stricter permissions. The dependency is that you are chasing a moving target.

Falsifiable test: pin a nightly, run your harness suite, then move to the next nightly and re-run. Count what broke. That number is the real cost of tracking upstream.