MCP Memory Gateway

What IgorGanapolsky/mcp-memory-gateway MCP server does

IgorGanapolsky/mcp-memory-gateway MCP server, presented in the repository as ThumbGate, places a check between an AI coding agent's proposed action and tool execution. Its primary use is preventing repeat mistakes during agentic development rather than retrieving general-purpose memory.

The default policy blocks detected secret leaks and two self-disable command classes: terminating the gate process and enabling the bypass environment override. Other listed risk patterns, including recursive deletion, force-pushes, fetch-and-run commands, and direct guardrail edits, produce a warning and log entry by default. Strict enforcement changes those warnings into denials through THUMBGATE_STRICT_ENFORCEMENT=1.

Feedback is saved as local lessons. A repeated, concrete failure can be promoted into a prevention rule, allowing later matching actions to surface the prior guidance or stop execution. The system operates at runtime and does not retrain or update model weights.

How it works

The enforcement path is local and does not require a cloud service. For agents with native hook support, ThumbGate runs in a pre-tool boundary such as PreToolUse or pre_tool_use. For MCP-compatible clients, it exposes the gate_check capability through a stdio MCP process. Cursor, Cline, OpenCode, and other MCP agents can use the advisory MCP path; the README describes direct hook enforcement for Claude Code, Codex, Gemini CLI, and ForgeCode.

A typical flow is to initialize the agent hooks, verify the wiring with doctor, and then capture feedback describing what happened and what should change. Later tool calls are matched against that stored context. The dashboard and evaluation commands provide local views and proof reports, but an empty dashboard after initial setup is considered a valid result.

The IgorGanapolsky/mcp-memory-gateway MCP server is therefore suited to teams that want controls applied before a tool call, rather than relying only on instructions in an agent prompt. The repository explicitly distinguishes a runtime hook from prompt advice: the hook can enforce the decision boundary before execution.

Setup and configuration

The README identifies the stdio MCP command as:

npx -y thumbgate serve

npx -y thumbgate mcp is documented as an equivalent command. Do not substitute npm start for MCP setup; that starts the hosted HTTP API instead of the agent-facing stdio process.

For the broader local setup, initialize the desired agent with npx thumbgate init, then run npx thumbgate doctor to confirm that the PreToolUse hook is installed. The repository also documents agent-specific initialization for Claude Code, Codex, Gemini CLI, ForgeCode, Cursor, Cline, and OpenCode. The enforcement mode can be tightened with THUMBGATE_STRICT_ENFORCEMENT=1; without it, the listed high-risk classes warn and log rather than deny.

Tools and capabilities

The IgorGanapolsky/mcp-memory-gateway MCP server supports the following documented operations and behaviors:

• Evaluate proposed tool calls before execution.
• Expose an MCP gate_check path over stdio.
• Hard-block detected secret leaks and direct self-disable attempts by default.
• Warn and log selected risky commands, including rm -rf and force-pushes.
• Capture thumbs-up or thumbs-down feedback with context and corrective guidance.
• Promote repeated failures into local prevention rules.
• Run prompt evaluations that produce reusable eval cases and local proof reports.
• Open a local dashboard for gate activity and statistics.

These capabilities are focused on coding-agent governance. The provided material does not describe a remote hosted MCP endpoint, a database-backed memory service, or model-training integration.