Black_Wall framework integrations

Drop-in guards that put a pre-action risk check in front of any tool your AI agent is about to run — send an email, run SQL, make a payment, delete a file. The agent's proposed action is sent to Black_Wall first; you get back a verdict, and the guard decides whether to let it through.

Black_Wall is a pre-action risk-check API for AI agents — a BLUETIER product. Free tier: ~100 forecasts/month, no credit card. Get a key at https://blackwalltier.com.

The decision model

Every forecast returns two things the guards key on:

Field	Values	Meaning
gate	AUTO · CONFIRM · HUMAN_REQUIRED	The actionable control — what should happen to the action
recommendation	GO · CAUTION · STOP	The human-readable verdict

The gate is derived from risk score (0–100) and reversibility (can it be undone?) — an irreversible action gets held at a lower risk than a reversible one. These guards map the gate to behaviour:

• AUTO → run the tool.
• CONFIRM → hold for a human (default: block; pass an on_confirm hook to allow).
• HUMAN_REQUIRED → block, return the red flags + safer alternatives.

If the API call fails or times out, the guards fail closed (block) by default — it's a safety layer, so the safe failure is to stop. Every guard exposes a flag to fail open instead if you'd rather the agent keep moving when the check is unavailable.

The endpoint

POST https://blackwalltier.com/api/v1/forecast
Authorization: Bearer bw_live_...
Content-Type: application/json

{
  "action": "send_email",
  "inputs": { "to": "client@acme.com", "subject": "...", "body": "..." },
  "context": { "agent_role": "support bot", "user_intent": "reply to a ticket" },
  "options": { "depth": "standard" }   // "deep" adds reasoning trace + mitigations
}

Response (the fields the guards use):

{
  "id": "fc_...",
  "recommendation": "CAUTION",
  "risk_score": 62,
  "gate": "CONFIRM",
  "reversibility": { "class": "RECOVERABLE", "rollback_cost": 50 },
  "red_flags": [{ "severity": "high", "code": "MISSING_AUTH", "message": "..." }],
  "alternative_actions": ["Ask the customer to confirm before refunding"],
  "tokens_charged": 87,
  "latency_ms": 3400
}

Latency is a few seconds (typically 4–8s standard, 10–13s deep) — this runs once before a consequential action, not on every token. There are 28 red-flag codes in the taxonomy.

Pick your framework

Framework	File
LangChain (Python)	langchain/
CrewAI (Python)	crewai/
Vercel AI SDK (TypeScript)	vercel-ai-sdk/
OpenAI tool calling (Python + TS)	openai/
Pydantic AI (Python)	pydantic-ai/
AutoGen / AG2 (Python)	autogen/
LlamaIndex (Python)	llamaindex/
n8n (no-code — HTTP Request + IF nodes)	n8n/
Stripe Agent Toolkit — gate money-moving actions (Python)	stripe-agent-toolkit/
PayPal Agent Toolkit — gate money-moving actions (Python)	paypal-agent-toolkit/
Coinbase AgentKit — gate on-chain (irreversible) actions (Python)	coinbase-agentkit/
Shopify — gate commerce actions (refunds, prices, orders) (Python)	shopify/
Twilio — gate SMS / voice / WhatsApp sends (Python)	twilio/
Cloud / infra — gate destructive AWS / Terraform / kubectl ops (Python + MCP)	cloud-infra/
LiteLLM Proxy — gateway guardrail (Python)	litellm/
MCP hosts — Cursor, Claude Code/Desktop, Windsurf, Antigravity	mcp/
Coding agents (Aider, Cline, OpenHands, Goose) — via MCP	coding-agents/

The framework files above are self-contained — copy one into your project, set BLACKWALL_API_KEY, and wrap your tools. No SDK dependency; just an HTTP call.

MCP hosts — Cursor, Claude Code/Desktop, Windsurf, Google Antigravity — can skip the code entirely: add the published blackwall-mcp server and the agent gets a forecast tool. Setup for each: mcp/ (Antigravity deep-dive: antigravity/).

Pairs with other tools

Some tools feed Black_Wall rather than wrap it. These ride in as context.prior_findings — offline risk analysis the runtime gate weights as priors.

Tool	What it adds	File
swarm-test — multi-agent reliability testing	Flags risky agents / tools / interaction edges offline; the gate weights them at runtime	swarm-test/