MCP Memory Gateway

What IgorGanapolsky/mcp-memory-gateway MCP server does

ThumbGate places a local decision point between an AI coding agent’s proposed action and the tool execution step. It is intended for agents that may otherwise delete files, expose secrets, force-push changes, fetch and run remote content, or edit their own guardrails. The gate evaluates the proposed call and returns an allow, warning, or denial outcome.

The default policy blocks detected secret leaks and two self-disable command classes: terminating the gate process and enabling its bypass environment override. Destructive or high-risk patterns such as rm -rf, force-pushes, fetch-and-run commands, and direct guardrail edits warn and log by default. Setting THUMBGATE_STRICT_ENFORCEMENT=1 changes those warnings into hard denials.

Feedback becomes local operational memory. A captured thumbs-down can record what went wrong and what should change; repeated concrete failures can then become prevention rules. The process does not retrain the model or update model weights.

How it works

The enforcement path runs locally and does not require a cloud service. Agent integrations can install hooks that inspect a tool call before it executes. Claude Code, Codex, and Gemini CLI use hard pre-action enforcement paths; Cursor, Cline, and other MCP-based setups use the gate_check MCP capability as an advisory check.

The project recommends wiring the hook first, verifying it, and only then adding matching feedback. npx thumbgate doctor confirms that the PreToolUse hook is installed, while the local dashboard can show activity. An empty dashboard is treated as a valid result during initial setup.

Setup and configuration

Run the initializer with npx thumbgate init, then verify the wiring with npx thumbgate doctor. To open the local dashboard, use npx thumbgate dashboard --open. Feedback can be captured from the CLI, for example:

npx thumbgate capture --feedback=down \
  --context="Never run DROP on production tables" \
  --what-went-wrong="agent proposed DROP" \
  --what-to-change="require review for DROP"

For MCP directory or client installation, start the stdio server with npx -y thumbgate serve; npx -y thumbgate mcp is equivalent. Do not substitute npm start, which launches the hosted HTTP API rather than the agent-facing stdio server.

Tools and capabilities

• Check proposed MCP tool calls with gate_check.
• Capture explicit thumbs-up or thumbs-down feedback as local lessons.
• Promote repeated concrete failures into prevention rules.
• Apply default hard blocks for secret leaks and gate self-disable attempts.
• Warn and log selected high-risk command classes, or deny them in strict mode.
• Generate reusable prompt-evaluation cases and local proof reports with npx thumbgate eval.
• Inspect local activity through the dashboard.

Limitations and notes

MCP clients that use gate_check receive advisory behavior; hard enforcement depends on an agent-specific pre-action hook being installed and firing. The default policy does not block every risky command: several high-risk classes warn and log unless strict enforcement is enabled. The enforcement path is local, and the system governs runtime tool calls rather than changing model behavior or weights.