Mcpcut

What mcpcut does

The mcpcut MCP server adds an inspection and control layer between an AI client and MCP servers. It records tool requests, responses, and policy decisions in a local journal, redacting secrets before data is written. The journal uses a hash chain, and the head can be signed and checked offline with a public key and an exported report.

Policies can allow, deny, or require approval for calls. Rules may target a server, tool, or tool class, with read-only tools able to pass while other operations wait. A held request remains pending until it is approved through the command line, web UI, terminal console, or an in-session confirmation where configured. If the agent stops waiting, the request is closed without sending the call.

The server also supports tool quarantine. New tools, or tools whose descriptions or schemas changed after being trusted, can be held for review with a change diff. For managed deployments, it provides agent identities, per-tool grants, groups, an encrypted credential vault, and one endpoint for the servers assigned to an agent.

How it works

The mcpcut MCP server can run in front of one existing server with wrap, requiring no initial identity or registry setup. The adopt command can find MCP configurations in Claude Code, Cursor, and Claude Desktop, show the proposed changes, and apply them while keeping backups for undo. The client then launches mcpcut as the intermediary.

For a broader setup, connect and serve use named servers from a registry, agent keys, and vault-held credentials. A pool endpoint at /mcp can expose all servers granted to an agent, while connect --url bridges a local stdio client to a remote service. The request path applies grants, policy, quarantine, and approval before forwarding the call.

Setup and configuration

Node.js 24 or newer is required; no separate dependency installation is needed for the quick-start flow. Run mcpcut through npx, or install it globally with npm for the setup wizard and managed service features.

A shared policy can be stored at ~/.mcpcut/data/policy.json. A manually wrapped server can instead receive a policy path immediately after wrap. The ui command opens the administrative interface, where policies can be created and per-tool actions can be set to client confirmation, approval, or denial.

On Windows, npm commands need to be launched through cmd /c; the adoption flow adds this form automatically. The initial setup can begin with a single wrapped server and expand to a registry, agent grants, credentials, and a service endpoint.

Tools and capabilities

• Wrap one MCP server or adopt existing client configurations.
• List sessions and inspect recorded requests and responses.
• Approve pending calls or manage them through the web and terminal interfaces.
• Apply per-server, per-tool, and tool-class policies.
• Quarantine new or modified tools for review.
• Manage declared agent folders with read, write, and edit grants; deleted files go to restorable trash.
• Generate signed audit material and verify exported reports offline.

The mcpcut MCP server can optionally use local Postgres for a complete file audit and local meaning-based search, according to the repository documentation. Administrative changes are attributed to named users with owner, operator, or viewer roles.