v1.18.0
takahirom/arbigentv1.18.0Aug 8, 2026by yamadashy
Upgrading?
AI Summary
Introduces a sandbox mode for the MCP server and implements security hardening to prevent unauthorized access and credential leakage.
Key Highlights
- New `--sandbox` flag to confine MCP server to a workspace directory
- File system tools restricted to sandbox mode only
- Credentials in repository URLs are redacted before logging
- Cloud metadata endpoints are refused as clone targets
Breaking Changes
- The `file_system_read_file` and `file_system_read_directory` tools are now sandbox-only
New Features
- MCP Sandbox Mode with virtualized results
- Security redaction for URLs in logs
- Blocking of cloud metadata endpoints
Full Release Notes
This release introduces sandbox mode for the MCP server, confining it to a single workspace directory, along with a set of security hardening changes across the MCP tool surface and remote repository processing! ## What's New 🚀 ### MCP Sandbox Mode (#1753, #1754) The MCP server can now be confined to a single workspace directory with the `--sandbox` flag: ```bash # Confine to the current working directory repomix --mcp --sandbox # Confine to a specific directory repomix --mcp --sandbox path/to/project ``` By default the MCP server can read any path the host user can, which is convenient for a trusted local assistant but too broad when the server is exposed to an untrusted client or agent, such as when embedding Repomix in a hosted application. In sandbox mode: - Every path is relative to the workspace root. Absolute paths, `~`, `..`, and Windows drive/UNC paths are refused, and paths that resolve outside the root (including through symlinks) are dropped. - Results and error messages are virtualized, so host paths are never exposed. - Only the read-only, root-confined tools are registered: `pack_codebase`, `read_repomix_output`, `grep_repomix_output`, `file_system_read_file`, and `file_system_read_directory`. Remote packing, skill generation, and attaching external outputs are disabled. This is an application-level confinement of the tool surface, not an OS-level sandbox: when hosting the server for untrusted clients, still run it under your platform's usual isolation (containers, dedicated users). See [MCP Server – Sandbox Mode](https://repomix.com/guide/mcp-server#sandbox-mode) for details. Special thanks to @huy-trn for designing and implementing this feature. Great to have a long-time contributor back! 🎉 ## Security Hardening 🔒 ### MCP File System Tools Are Now Sandbox-Only The `file_system_read_file` and `file_system_read_directory` tools are now registered **only in sandbox mode**. Outside `--sandbox` they could read any path the process can, while their built-in Secretlint scan was presented as more protection than it provides: it recognizes known secret formats in file content, and nothing else. In sandbox mode the workspace root gives these tools a real access boundary, with the secret scan kept as an additional heuristic safeguard. If your MCP setup relied on these two tools, add `--sandbox [dir]` to the server arguments to keep using them, confined to that workspace. Tool descriptions and documentation have also been reworded across the board to describe the Secretlint scan as what it is: a best-effort content heuristic, not an access boundary. See the security advisory for details. Thanks to @rafaelfiguereod-stack for the report! 🙏 https://github.com/yamadashy/repomix/security/advisories/GHSA-rpmv-562j-qxrv ### Credentials in Repository URLs Are Redacted Before Logging Remote URLs are routinely written with credentials inline (`https://<token>@github.com/owner/repo.git`), especially in CI. Repomix previously echoed such URLs as-is into console output, trace logs, error messages (git repeats the full command line in its failure output), and the MCP `pack_remote_repository` response, persisting the credential in terminal scrollback, CI build logs, and log aggregation systems. All of these paths now redact userinfo credentials and known credential query parameters (`token`, `private_token`, `access_token`, and similar) before the URL is written anywhere. See the security advisory for details. Thanks to @kakashi-kx for the report! 🙏 https://github.com/yamadashy/repomix/security/advisories/GHSA-w8cw-mgw9-74h7 ### Cloud Metadata Endpoints Are Refused as Clone Targets Remote packing now refuses to clone from cloud instance metadata endpoints (`169.254.0.0/16`, `fd00:ec2::254`, `100.100.100.200`, `metadata.google.internal`), which serve credentials and host no git repository. Through the MCP server an AI agent chooses the URL, so an injected instruction could otherwise aim a clone at an address the user never asked for. Private networks (RFC1918) remain allowed, since cloning from self-hosted git servers is normal. ## How to Update ```bash npm update -g repomix ``` --- As always, if you have any issues or suggestions, please let us know on GitHub issues or our [Discord community](https://discord.gg/wNYzTwZFku).
Related Repositories
BreezeAPP 是一款為 Android 和 iOS 平台開發的純手機 AI 應用程式。從 App Store下載,即可在不連網的狀態下享受多項 AI 功能。源碼由聯發創新基地(MediaTek Research)提供。我們旨在推廣兩個概念: 人人都可以在自己的手機上自由選擇並運行不同的LLM - one is free to choose one's own LLM to run on a phone,以及任何app開發者都可以輕鬆寫作創意的純手機AI應用 - any dev can create purely phone-based AI apps easily。
A collective list of free APIs
The AI that really does things. Any OS. Any Platform. The lobster way. 🦞
Interactive roadmaps, guides and other educational content to help developers grow in their careers.
Reverse engineer anything with agents, from app behavior down to native binaries.