Install in 30 seconds
Needs Claude Code and Node.js 22.13+. No dependencies. In a Claude Code session:
/plugin marketplace add gabe4coding/claude-code-mode
/plugin install code-mode@claude-code-mode
Start a new session and ask for anything that needs several MCP calls. Using auto mode with Claude Code older than 2.1.301? Add allow rules first: Permissions.
See it in 43 seconds
The full film with sound:
https://github.com/user-attachments/assets/468c412f-ecae-4e04-a53b-8a291561a052
| Without code-mode | With code-mode |
|---|
| Claude calls tools one by one | Claude writes one program |
| Every full result lands in context | Results are filtered in a sandbox |
| Context fills with raw JSON | Only the answer comes back |
Why
- Fewer tokens. Intermediate data never enters the context. See Measured.
- Sandboxed. Separate Node process, no file system, CPU limit. No network on macOS and on Linux where
unshare -rn works.
- Learns your tools. Hints per MCP server (formats, required args, limits). Claude proposes fixes. You approve.
- Reusable functions. Programs that work can become saved functions for later sessions, after you read and approve the code.
- Your permissions still apply. Every MCP call inside the program gets the normal permission check.
Measured
The bench in bench/ runs six tasks in headless Claude Code against a fake MCP server.
Each task has traps: pages, nested fields, bots and results larger than 20,000 characters.
Haiku 5.5, without and with code-mode (5 runs of each task, plugin 0.6.0 and 0.6.1, 2026-10-10).
In 0.6.0 the code-mode start line did not reach the model, so Haiku called most tools directly.
| Tools called directly (0.6.0) | Through code-mode (0.6.1) |
|---|
| Share of MCP calls made directly | 89% | 2% |
| Correct answers | 30/30 | 30/30 |
| Cost per task | $0.007 | $0.004 |
| Input tokens per task | 167,297 | 149,995 |
| Output tokens per task | 1,814 | 1,194 |
| Seconds per task | 13.5 | 11.6 |
Opus 5.5, output projection off and on: duplicate MCP calls per task went from 1.37 to 0.10.
Cost and correct answers did not change. Refer to PR #11.
Limits: these are bench tasks in headless runs. Interactive sessions were not measured.
Most input tokens are the fixed context of Claude Code, so the input saving is smaller than the cost saving.
To run the bench again, refer to Development.
Example
"Which of my open Jira issues have no update in the last 14 days?"
const issues = await call("mcp__claude_ai_Jira__searchJiraIssuesUsingJql", {
cloudId: "example.atlassian.net",
jql: "assignee = currentUser() AND statusCategory != Done",
})
const cutoff = Date.now() - 14 * 24 * 3600 * 1000
return issues.issues
.filter(i => Date.parse(i.fields.updated) < cutoff)
.map(i => ({ key: i.key, title: i.fields.summary }))
Claude sees a short list of keys and titles, not every full issue.
Documentation
| Page | |
|---|
| Installation | Requirements, install, local folder, update |
| Usage | The three tools, how to write a program, the result |
| Hints | Write hints, approve proposals, the safety rules |
| Saved functions | Approve or write functions that programs use again |
| Permissions | Manual and auto mode, allow rules |
| Configuration | Every option and its default |
| How it works | One run from start to end, the four sandbox layers |
| Troubleshooting | Common errors and known limits |
| Development | Layout, tests, releases |
Contributing / feedback
Issues and PRs welcome. If code-mode saves you tokens, a ⭐ helps others find it.
License
MIT. See LICENSE.
