mcp-toolbox
Use when working with or extending the local Codex MCP toolset/server that exposes HTTP endpoints for shell, filesystem, git, and Excel operations, or when a task should be executed by calling those MCP endpoints instead of local tools, especially if the MCP server is running.
When & Why to Use This Skill
The mcp-toolbox skill serves as a comprehensive interface for the Codex MCP (Model Context Protocol) server, enabling Claude to interact with local development environments through standardized HTTP endpoints. It provides a secure and structured way to execute shell commands, manage file systems, perform Git operations, and edit Excel files, prioritizing remote server execution over local tools to ensure workflow consistency and safety.
Use Cases
- Automated Git Management: Use the skill to check repository status, review code diffs, and perform commits through the MCP Git endpoints, streamlining version control workflows.
- Secure File Operations: Read and write files within a restricted base directory, allowing the agent to manage project assets while adhering to defined security boundaries.
- Programmatic Excel Editing: Execute complex spreadsheet updates using a multi-step 'preview-then-commit' workflow to validate changes and maintain data integrity.
- Environment-Agnostic Shell Execution: Run system commands and scripts via the /shell/exec endpoint, facilitating task automation without relying on direct local shell access.
- MCP Server Maintenance: Extend or debug the MCP toolset by editing tool modules and updating server configurations directly through the toolbox interface.
| name | mcp-toolbox |
|---|---|
| description | Use when working with or extending the local Codex MCP toolset/server that exposes HTTP endpoints for shell, filesystem, git, and Excel operations, or when a task should be executed by calling those MCP endpoints instead of local tools, especially if the MCP server is running. |
MCP Toolbox
Prefer MCP endpoints when they can satisfy the request and the server is reachable.
Quick start
- Ask whether the MCP server is running if it is not already clear.
- If uncertain, do a lightweight reachability check (for example, a /shell/exec call with a harmless command) before assuming the server is unavailable.
- Use endpoint details and payload shapes from references/toolbox.md.
Guided workflow
- Confirm the server base URL and whether it is running.
- Call /health to verify reachability; if it fails, fall back to local tools or ask to start the server.
- Use /tools/list or /openapi.json to discover tool schemas for auto-invocation.
- Use the smallest endpoint that satisfies the request; avoid local shell when /shell/exec can do it.
- Keep file paths within MCP_ALLOWED_BASE; ask for a relative path if unsure.
- Follow inspect -> read_range -> preview_write -> commit_write for Excel edits.
- Show /git/status or /git/diff before /git/commit and remind that commit stages all changes.
Using the endpoints
- Route shell commands through /shell/exec instead of local execution when the server is live.
- Use /fs/read and /fs/write for file operations; respect the allowed base directory enforced by the server.
- Use /git/status and /git/diff before /git/commit; note that /git/commit stages all changes.
- For Excel edits, always call /excel/preview_write before /excel/commit_write to validate changes.
Maintaining or extending the server
- Edit the tool modules in server/tools/ and wire them into server/mcp_server.py.
- Keep path resolution and safety checks aligned with server/config.py (ALLOWED_BASE_DIR and MAX_READ_BYTES).
- Update README.md if endpoints, behavior, or setup steps change.
- When adding dependencies, update server/requirements.txt.