ProceedGate — Stop Scraping Storms Before They Burn Your Budget

Without Guardrails, This Happens

Every team running scraping or AI agents hits these — usually at 3 AM.

🌀

Retry Storms

One failed API call triggers 500 retries in under a minute. Loops burn budget in minutes.

Typical damage: $50–$500 per incident

💸

Invisible Cost Leaks

No guardrails means wasted credits. Dashboards show the spend after it happened.

Average detection: 4+ hours too late

🚫

Late Alerts

Monitoring tools show you the fire. ProceedGate stops it before it starts.

Dashboards show — they don't stop

→ Start in Shadow Mode — see what would be blocked

Real Output

This Is What ProceedGate Actually Returns

Not mockups. Real API response structures from production.

Agent:    my-agent
Pattern:  47 identical requests / 60s
Zone:     RED (storm)
Action:   BLOCKED
Credits saved: 142

Loop detection:
  pattern_hash:    sha256:a3f8c2...
  count_in_window: 47 (threshold: 10)
  avg_interval_ms: 1274
  interval_cv:     0.08 (mechanical)
  backoff_detected: false

Webhook → Slack #alerts ✓

$ curl governor.proceedgate.dev/v1/governor/check

{
  "decision": "proceed",
  "zone": "safe",
  "credits_remaining": 1847,
  "loop_detected": false,
  "pattern_count": 2,
  "proceed_token": "eyJhbGciOiJFUzI1NiIs..."
}

Integration

Three Steps. Full Control.

1

Install the SDK

git clone https://github.com/loquit-doru/agent-cost-governor

2

Wrap Your Agent Calls

Add governor.gate() before expensive operations

3

Set Budget & Deploy

Per-agent limits, per-task caps — AI handles the rest

        agent.ts
      

import { CostGovernor } from './sdk-node/src/index';

const gov = new CostGovernor({ apiKey: process.env.PG_KEY });

const { decision } = await gov.gate({
  agent: 'scraper-v2',
  action: 'api_call',
  cost: 0.03,
});

if (decision === 'proceed') {
  await callExpensiveAPI();
}
      

→ Get your API key in 30 seconds

Try It Now

See It In Your Terminal

One command. 30 seconds. See what ProceedGate would block in your workflow.

## Clone and run a quick check
$ git clone https://github.com/loquit-doru/agent-cost-governor
$ cd agent-cost-governor && npm install
$ npx tsx runner/src/index.ts check --workspace demo

Scanning agent: scraper-v2...
Requests analyzed:  147
Loops detected:    3
Estimated savings: $42.80
Shadow blocks:     12 (would have blocked)

✓ Report saved to ./proceedgate-report.json

Shadow Mode · No enforcement risk · See results in 30s

Detection

3-Zone Detection + AI Reasoning

Pattern heuristics for speed, Workers AI (Llama 3.1 8B) for ambiguous cases — explainable, auditable, open source.

🔢

Pattern Counting

SHA-256 hashed request patterns. 10+ identical requests/min = storm detected.

🚦

3-Zone Classification

Green (safe), Yellow (gray zone), Red (storm) — graduated response per zone.

🧠

Gray Zone AI

Workers AI (Llama 3.1 8B) reasons over ambiguous 6–10 req/min cases. Full reasoning chains.

⏱️

Timing Analysis

Interval coefficient of variation + backoff detection. Catches mechanical vs human patterns.

💰

Budget Enforcement

Per-workspace credit limits. Hard gate at zero balance — no silent overruns.

🔐

Proceed Tokens

ES256 signed JWTs with JWKS verification. Cryptographic proof of authorization.

🔔

Webhook Alerts

Real-time Slack/Discord notifications on blocks. HMAC-signed for security.

📊

Cost Tracking

Per-agent usage analytics. See exactly how much each agent spends and saves.

🗺️ Roadmap: Adaptive baselines (EWMA+CUSUM), behavior fingerprinting, cross-workspace intelligence, decision caching.

→ Start free — protect your first 5,000 events

Use Cases

Built For Teams That Can't Afford Runaway Costs

🕷️

Scraping Agents

Apify, SerpAPI, Firecrawl, Bright Data — retry storms are the #1 cost killer.

Deep dive →

🤖

MCP & LLM Pipelines

LangChain, AutoGPT, CrewAI — any pipeline calling external APIs can silently loop.

⚙️

SaaS Automations

Background AI workflows — email, content, data processing — that can fail-loop silently.

🏢

Agencies & Multi-Tenant

Per-client budgets, per-agent limits. Hundreds of agents, zero rogue tasks.

Designed for teams building:

Web Crawling SEO Monitoring Data Extraction MCP Backends AI Pipelines DeFi Agents on BNB Chain

→ See all use cases with live demos

🛡️ Shadow Mode

Not Ready to Block? Just Watch.

Start in Shadow Mode. See what ProceedGate would block — without touching your running agents. Switch to enforcement when you trust the data.

Full visibility into retry patterns
Zero impact on running agents
Switch to enforcement when confident
Real data before real decisions

Try Shadow Mode Free

          agent.ts
        

const gov = new CostGovernor({
  apiKey: process.env.PG_KEY,
  mode: 'LOG_ONLY',  // shadow mode
});

const result = await gov.gate({
  agent: 'scraper-v2',
  action: 'api_call',
  cost: 0.03,
});

// Always proceeds, but logs
// what it would block
if (result.would_block) {
  console.warn('Storm detected');
}
        

On-Chain

Verified on BNB Chain

Real on-chain payment verification via x402 protocol. Agents pay micro-fees on BSC to unlock blocked decisions — cryptographic proof, not promises.

⛓️

BSC Payment Verification

x402-style paywall: agents submit BSC transaction hashes to unlock blocked decisions. Verified on-chain.

📜

Smart Contract (BSC Testnet)

Deployed & verified on Sourcify. Transparent, auditable governance on-chain.

🔐

ES256 Proceed Tokens

After payment, agents receive cryptographically signed JWTs. Verified via JWKS — tamper-proof authorization.

View BSC Mainnet Proof TX ↗

What ProceedGate Is NOT

✗ An observability dashboard that shows you the fire after it burns

✗ A wrapper around OpenAI or any single provider

✗ A vendor lock-in proxy — it's fully open source

✗ A dumb rate limiter — it uses AI reasoning for ambiguous cases

🧪 Early Access — Built in Public

Start Free. Scale When Ready.

👤

Direct access to founder

🗺️

Influence the roadmap

🔓

No lock-in. Cancel anytime.

Start Free Trial 14 days · Full features · No credit card Read the Docs

Frequently Asked Questions

Can't find your answer? Reach out — we reply fast.

How does ProceedGate detect retry storms?

ProceedGate hashes each request pattern (SHA-256) and counts identical patterns per minute. If more than 10 identical requests arrive within 60 seconds, the system escalates to the red zone and blocks. For ambiguous cases (6–10 req/min), Workers AI (Llama 3.1 8B) analyzes timing patterns, interval regularity, and backoff behavior to make the final allow/block decision with full reasoning chains.

Will it slow down my agents?

ProceedGate runs on Cloudflare's edge network. Simple allow/block checks are fast (pattern lookup in Durable Objects). Gray zone checks that invoke Workers AI add more latency but only trigger for ambiguous cases (6–10 req/min). A typical external API call takes 500–3000ms, so the overhead is modest relative to the operations being protected.

What is Shadow Mode (LOG_ONLY)?

Shadow Mode (LOG_ONLY) lets you observe what ProceedGate would block — without actually blocking anything. Your agents keep running normally while you get full visibility into patterns. Switch to enforcement when you trust the data. It's the best way to start: zero risk, real insights.

What happens when an agent hits its budget?

The gate returns a blocked response with a clear signal. You get a webhook notification (Slack, Discord, any URL). You can manually approve more spend via the dashboard — no runaway charges.

Which frameworks are supported?

Any framework that can make HTTP calls. Node.js SDK for native integration, MCP server for AI-native tool access. Works with LangChain, AutoGPT, Apify, CrewAI, and custom agents. REST API for non-Node environments.

Is this open source?

Yes. Full source on GitHub — SDK, worker, MCP server, everything. Inspect the code, contribute, or self-host.

How does BNB Chain integration work?

ProceedGate uses the x402 payment protocol on BSC (BNB Smart Chain). When an agent hits a cost gate, it can submit a BSC transaction hash to unlock the decision. The payment is verified on-chain — no trust required. We also have a verified smart contract on BSC Testnet for governance operations. View proof TX on BscScan →