Claude Work MCP

Local MCP server for connecting Claude Code to your work environment through explicit, auditable tools.

What it exposes

• list_work_tools: shows the active config, workspace root, links, endpoints, and safety gates.
• list_work_files: lists files under workspaceRoot.
• read_work_file: reads UTF-8 text files under workspaceRoot.
• search_work_files: searches text files under workspaceRoot.
• write_work_file: writes files only when allowWrite=true.
• read_handoff / append_handoff: shared Codex-Claude notes.
• call_work_api: calls only HTTP endpoints defined in config/work-tools.local.json.
• run_work_command: runs only allowlisted commands and only when allowCommands=true.

Install dependencies

cd "C:\Users\Peruano Pinto\Documents\Codex\2026-07-09\necesito\work\claude-work-mcp"
npm install
npm run build
npm run smoke

Register with Claude Code on Windows

After installing Claude Code, run:

cd "C:\Users\Peruano Pinto\Documents\Codex\2026-07-09\necesito\work\claude-work-mcp"
.\scripts\register-claude.ps1

Manual equivalent:

claude mcp add --scope user --transport stdio --env "WORK_MCP_CONFIG=C:\Users\Peruano Pinto\Documents\Codex\2026-07-09\necesito\work\claude-work-mcp\config\work-tools.local.json" claude-work-mcp -- node "C:\Users\Peruano Pinto\Documents\Codex\2026-07-09\necesito\work\claude-work-mcp\dist\server.js"
claude mcp list

If you prefer JSON config, copy claude-code.mcp.json.example into a Claude Code project as .mcp.json, then start Claude Code and approve the project MCP server.

Register with Claude Code on Linux/macOS

cd /path/to/claude-work-mcp
npm install
npm run build
./scripts/register-claude.sh

Configure your tools

Edit config/work-tools.local.json.

Add a direct API endpoint:

{
  "id": "crm_lookup",
  "name": "CRM contact lookup",
  "method": "GET",
  "url": "https://api.example.com/contacts?email=${params.email}",
  "headers": {
    "Authorization": "Bearer ${env:CRM_API_TOKEN}"
  }
}

Then set the token before starting Claude Code:

$env:CRM_API_TOKEN = "..."

Command execution and file writes are off by default. Turn them on only when you want Claude Code to make changes:

{
  "allowWrite": true,
  "allowCommands": true,
  "commandAllowlist": [{ "command": "git" }, { "command": "npm" }]
}

Suggested Claude prompt

Use the claude-work-mcp MCP server. First call list_work_tools, then read_handoff. Work only inside the configured workspace. If you need to change files or run commands and the tool is disabled, tell me exactly what config change you need.

Safety model

This server does not call arbitrary URLs. It only calls endpoints listed in config.

It does not access arbitrary filesystem paths. All workspace file tools stay inside workspaceRoot.

It does not run shell commands unless allowCommands=true, and even then only command names in commandAllowlist. Interpreter commands (node, py, bash, etc.) are additionally blocked from running with inline-eval flags (-e, -c, --eval, ...) even when allowlisted, since those flags bypass the workspaceRoot sandbox.

Single-file reads and writes also resolve symlinks and reject any target whose real path falls outside workspaceRoot, so a symlink planted inside the workspace cannot be used to read or write files elsewhere on disk.