BLACK_WALL — Pre-action risk gate for AI agents

What is BLACK_WALL?

BLACK_WALL is a pre-action risk gate for AI agents. It exposes a single tool, forecast, that the agent calls before any irreversible action such as sending an email, making a payment, running SQL, deleting data, or posting content.

How to use BLACK_WALL?

Configure BLACK_WALL in your MCP-compatible client (Claude Desktop, Claude Code, Cursor, or Windsurf) by adding a JSON block with the npx command and your API key. The agent then invokes the forecast tool before each potentially irreversible action; a normal email clears in one call, while a dangerous action returns a hard gate verdict.

Key features of BLACK_WALL

• Single forecast tool for pre-action risk assessment
• Risk score from 0–100 and reversibility class with rollback cost
• Gate verdict of proceed, confirm, or human-required
• 28 documented failure modes with named red flags (e.g. SQL_NO_WHERE, PROMPT_INJECTION_LIKELY)
• Hard gate that the agent cannot override
• Verdict in approximately 4–8 seconds

Use cases of BLACK_WALL

• Prevent accidental email sends to large recipient lists
• Stop SQL commands that lack a WHERE clause from executing
• Block payment or fund transfer actions that cross environments
• Intercept content posts that contain likely prompt injection attempts
• Require human approval for irreversible data deletion actions

FAQ from BLACK_WALL

What does BLACK_WALL do exactly?

BLACK_WALL acts as a brake for AI agents. The agent calls the forecast tool before any irreversible action, and the server returns a risk score, reversibility class, gate verdict, and named red flags — preventing the agent from talking itself into a bad action.

How do I install BLACK_WALL?

Add a JSON block to your MCP client configuration with command npx -y blackwall-mcp and environment variable BLACKWALL_API_KEY set to your API key.

Where do I get an API key?

Get a free API key at https://blackwalltier.com. The free tier includes approximately 100 forecasts per month and requires no credit card.

How fast is a verdict?

A typical verdict takes approximately 4–8 seconds.

Can the agent override a STOP verdict?

No. The verdict is a hard gate — the agent cannot talk itself into the bad action, because the gate is enforced by the tool's return value.