v0.10.1

code-mode is a Claude Code plugin. The model writes one JavaScript program that calls MCP tools in a sandbox, and only the return value goes back into the context.

Install

Type these commands at the prompt of a Claude Code session:

/plugin marketplace add gabe4coding/claude-code-mode
/plugin install code-mode@claude-code-mode

You need Claude Code and Node.js 22.13 or later. If you have the plugin, update it to get this version.

New in 0.10.1

• The run footer shows each MCP call that had no reply when the run ended, as - <tool> did not finish. Before, a timeout did not tell which tool hung.
• A hung MCP call no longer stops the result of run_code. When the sandbox ends, the plugin waits at most 1 second more for running calls.
• A call that waits in the queue does not start after the sandbox ends. (#20)

Changes since v0.3.0

• Saved functions (0.9.0): the model can propose a program as a function, and later programs call it with lib.<name>(). You approve each one.
• MCP resources (0.9.0): programs can list and read the resources of MCP servers.
• Network block on Linux (0.9.0): the sandbox has no network on Linux (unshare) and on macOS (sandbox-exec).
• Output projection (0.6.0): the footer shows the size and the shape of each MCP result, and recall(n) reads a result again with no new call.
• Hints (0.4.0 – 0.5.0, 0.10.0): the model proposes short notes about a server, and you approve or discard them in the band above the prompt. The model can also propose to remove a wrong hint.
• Files (0.8.0): saveFile and readFile, argument names learned from earlier calls, and a check of the program before each run.
• Limits (0.7.0): sleep(), at most 8 MCP calls at one time, and result shapes in search_tools.

Docs: README and docs.