@cipherrow/gateway

Run 8x more agents in the same context window. Official SDK for CR Gateway — validate every LLM output, kill bad swarm chains early, compress context by 70%+ per pass.

• 8x agent scaling on GPT-4o 128K (proven in live benchmarks)
• 100% precision in benchmark testing — zero false positives across 1,800+ test calls
• ~70% token savings per compression pass
• <5ms processing — no LLM calls, pure pattern matching
• Zero dependencies — works anywhere fetch is available: Node.js 18+, Deno, Bun, Cloudflare Workers

Install

npm install @cipherrow/gateway

Quick Start

import { CRGateway } from "@cipherrow/gateway";

const gw = new CRGateway("your-api-key");

// Validate any LLM response before acting on it
const result = await gw.validate({
  type: "assistant_response",
  content: "Based on the data, I recommend proceeding with Option A.",
});

if (result.valid) {
  // Safe to use
} else {
  // Failed validation — don't act on it
  console.log("Blocked:", result.checks);
}

Configuration

new CRGateway(apiKey: string, options?: {
  baseUrl?: string;    // default: CR Gateway production
  timeout?: number;    // request timeout in ms (default: 10000)
  maxRetries?: number; // retries for 429/5xx errors (default: 2)
})

API Methods

Validation

• validate(message, options?) — Check an LLM response for safety issues. Returns pass/fail.
• validateBatch(messages) — Validate up to 50 messages in one call.
• store(message, sessionId) — Validate and store a message.
• getMessage(messageId) — Retrieve a stored message.

Swarm Management

• swarmCheck(agentChain, options?) — Should this agent chain keep running or be killed? Saves wasted LLM calls on bad chains.

Context & Compression

• compress(messages, options?) — Reduce conversation size for cheaper LLM calls.
• contextCheck(params) — Should the agent save work, wrap up, or keep going?

Agent Relay

• relay(message, deliverTo, sessionId) — Validated agent-to-agent message delivery.

Orchestrator

• createWorkflow(params) / listWorkflows() / deleteWorkflow(id) — Manage workflow definitions.
• startRun(workflowId, input?) / getRun(id) / cancelRun(id) — Execute and monitor workflows.
• completeStep(runId, stepId, params) — Report step completion.

Configuration & Usage

• getConfig() / updateConfig(updates) — Read/update your tenant settings.
• getUsage(days?) — Usage statistics.
• feedback(params) — Report outcomes to improve accuracy over time.
• oracleStatus() — Check your validation model status.

Key Management

• rotateKey() — Generate a new API key.
• listKeys() — List all keys (masked).
• revokeKey(keyId) — Revoke a key.

Authentication

• register(params) / login(params) / logout(token) — Account management.
• checkSession(token) — Verify session.
• forgotPassword(params) / resetPassword(params) — Password recovery.

Error Handling

import { CRGateway, GatewayError } from "@cipherrow/gateway";

try {
  await gw.validate(message);
} catch (err) {
  if (err instanceof GatewayError) {
    console.error(err.status);     // HTTP status code
    console.error(err.requestId);  // for support
  }
}

429 and 5xx errors are automatically retried with exponential backoff.

Rate Limits

await gw.validate(message);
console.log(gw.lastRateLimit);
// { limit: 60, remaining: 57, retryAfter: null }

Documentation

Full docs: https://gateway.cipherandrow.com/docs.html

License

MIT