BLACK_WALL — Pre-action risk gate for AI agents

A guardrail for AI agents, as an MCP server.Your agent calls one tool —forecast— before any irreversible action (send email, move money, run SQL, delete data, post content). It gets back a risk score (0–100), a reversibility class, aGO/CAUTION/STOPrecommendation, and named red flags in a few seconds (~4-8s).

Works in any MCP host:Claude Desktop, Claude Code, Cursor, Windsurf, and any agent framework with MCP support.

The wall between your agent and disaster. A BLUETIER product.

Sign up free athttps://blackwalltier.com→ Dashboard → API keys → Create key. Free tier: ~100 forecasts/month, no card. Your key looks likebw_live_….

{ "mcpServers": { "blackwall": { "command": "npx", "args": ["-y", "blackwall-mcp"], "env": { "BLACKWALL_API_KEY": "bw_live_your_key_here" } } } }

Restart Claude Desktop. You'll see aforecasttool available.

Settings → MCP → Add new global MCP server, then inmcp.json:

{ "mcpServers": { "blackwall": { "command": "npx", "args": ["-y", "blackwall-mcp"], "env": { "BLACKWALL_API_KEY": "bw_live_your_key_here" } } } }

claude mcp add blackwall -e BLACKWALL_API_KEY=bw_live_your_key_here -- npx -y blackwall-mcp

BLACKWALL_API_KEY=bw_live_your_key_here npx -y blackwall-mcp

Once added, instruct your agent:"Before any irreversible action, call theforecasttool and stop if it returns STOP."The model will call it automatically when it's about to do something risky.

Returns:recommendation (GO/CAUTION/STOP),risk_score(0–100),reversibility(class + rollback cost),gate(proceed/confirm/human-required),confidence,red_flags[],predicted_result,alternative_actions[].

Agent about to runDELETE FROM users;(no WHERE clause) →

🛑 BLACK_WALL: STOP — risk 99/100 Red flags: • [CRITICAL] SQL_NO_WHERE — deletes the entire table, not one row • [CRITICAL] INTENT_MISMATCH — intent was "remove a single test row" • [CRITICAL] IRREVERSIBLE_NO_BACKUP — no recovery path Guidance: DO NOT take this action. Surface the red flags to the user.

Not ready to let a guardrail block your agents? Start inobserve mode. It scores and logs every action butnever tells the agent to stop— your agents behave exactly as they do today. After a week, review your dashboard and see what itwouldhave caught.

{ "mcpServers": { "blackwall": { "command": "npx", "args": ["-y", "blackwall-mcp"], "env": { "BLACKWALL_API_KEY": "bw_live_your_key_here", "BLACKWALL_MODE": "observe" } } } }

Then see"what your agents almost did"in your dashboard. FlipBLACKWALL_MODEtoenforce(or just remove it — enforce is the default) when you're ready to actually block.

- forecast— pre-action risk check. ReturnsGO/CAUTION/STOP, risk score, named red flags, reversibility class, and a verifiable receipt.
- observe— post-action outcome report. Tells BLACK_WALL what actually happened after the action ran (or after the agent obeyed a STOP verdict). Closes the loop so the system can track prediction accuracy over time. FREE — no tokens charged.

Wire your agent to callforecastbefore any irreversible action, then callobserveafterwards with theforecast_idfrom the original response.observeaccepts anoutcome_class(matched/over_scope/under_scope/no_op/diverged/aborted) and optionaldivergence_severityanddetails. See theforecastexample below; the same wiring applies toobserve.

Use it in code — thegate()control (any JS/TS agent)

Running an agent in Node (LangChain, a custom loop, ElizaOS, a cron job)? You don't need an MCP host — call BLACK_WALL straight from the library, and letgate()make the checkimpossible to skip. One wrap forecasts the action, enforces the verdict (fails closedonSTOP/ unknown / unreachable), runs your side effect only when allowed, and reports the real outcome withobserveautomatically.

import { gate, BlackWallBlocked } from 'blackwall-mcp/lib/gate'; // Wrap ANY risky action in a few lines. BLACKWALL_API_KEY lives in the env. try { const { result } = await gate( { action: 'run_sql', inputs: { statement: sql }, context: { user_intent } }, () => db.query(sql), // your real side effect — only runs if allowed { onCaution: (v) => confirmWithHuman(v) }, // CAUTION needs a yes; default = block ); // ...use result } catch (e) { if (e instanceof BlackWallBlocked) { // STOP, unconfirmed CAUTION, or forecast unavailable → the action NEVER ran console.error('Blocked:', e.reason, e.verdict?.red_flags); } else throw e; // a real error thrown by your action }

Fails closed by design.If no verdict can be obtained (network / auth / timeout), the action doesnotrun unless you explicitly passfailOpen: true. A risk gate that fails open is not a risk gate. The loop closes itself —gate()callsobservewith the actual outcome (matched/diverged/aborted), so your forecasts sharpen over time.

Prefer the lower-level pieces? They're exported too:

import { forecast, observe } from 'blackwall-mcp/lib'; const v = await forecast({ action: 'make_payment', inputs: { amount_usd: 50000 } }); if (v.recommendation === 'STOP') throw new Error('halt'); // ... take the action ... await observe(v.id, { outcome_class: 'matched' });

Runnable demo:examples/gate-quickstart.mjs.

- Published keys:https://blackwalltier.com/.well-known/blackwall-signing-keys.json(stable, cacheable)
- Stateless verify endpoint:POST https://blackwalltier.com/api/v1/receipts/verifywith{ envelope, request_body, response_body }
- Hashes only — BLACK_WALL never stores the raw request/response bodies, so receipts give cryptographic audit without payload exposure
- Free-tier retention: 90 days. Paid: indefinite.

- Site & docs:https://blackwalltier.com
- Get a key:https://blackwalltier.com/dashboard/keys

Trust and safety layer for AI agents — scores MCP servers for security risk, capability flags, and prompt injection.