Smart Behavioral Firewall

Don't let agents run wild.Use protection.

A smart behavioral firewall that learns how your agents normally behave, catches weird calls and prompt injection, and routes the risky ones through human approval.

Behavioral FirewallHuman-in-the-LoopAgent CredentialsEscrow Contracts

/ The problem

Agents can hit IAM, Stripe, Terraform, calendars, CRMs

That's the point.
The risk is when they do it with injected instructions, drift, or without a real trust layer.

Prompt injection & jailbreaks

Bad actors hijack tool calls

Drifted reasoning & loops

Agents stuck in infinite retry cycles

Overreach & wrong environment

Tools called without authorization

Unverified agent identity

No way to prove who initiated a call

/ How it works

It learns your agents

LetsPing builds a behavioral graph of how your agents move between tools and actions. Every step is scored. Anomalies surface before they become incidents.

Behavioral profilingMarkov anomaly scoringBaseline building

/ How it works

Guardrails from day one

Enterprise grade rules fire immediately on setup, no training required. PII is redacted in transit. Injection attacks are blocked. Egress is locked to your allowlist.

DLP redactionSSRF blockedEgress firewallShadow mode

/ How it works

Enforcement once it's sure

Once the baseline locks, the firewall flags unmapped sequences and routes them to your console. Humans approve. Agents learn. Nothing unexpected reaches production.

Real-time interceptsHITL triage queueAuto-block
Learn
Guard
Enforce

/ Human in the loop

Humans looped when needed

When the firewall pauses a call, operators see the full context, patch the payload, or reject it. The run then resumes safely.

Approvals train the system.
What you repeatedly approve becomes normal; what you reject stays suspicious.

Resolve from your console, CLI, or phone.

Connect your preferred channel. No app store required.

Pending Review
Critical
Service
cx-agent
Action
stripe:refund
Anomaly Score
0.91
Reason
Off-behavior transition
Payload Diff
"amount": 4800
+"amount": 480
 "currency": "usd"
 "reason": "duplicate_charge"

/ Agent identity

Agents with credentials.
Not shared secrets.

Give each agent its own ID and secret. Every call is signed; handoffs carry escrow envelopes so downstream services can verify who initiated what.

If you're a SaaS or storefront, you can accept agents as first-class citizens by verifying LetsPing webhooks + escrow.

Capabilitiesv0.3.3
  • Each agent gets its own ID & secret
  • Every call is cryptographically signed
  • Escrow envelopes for verified handoffs
  • Downstream services verify who initiated what
  • Trust score (0–999) others can verify
  • Payment mandates: x402 / AP2 compatible

/ Observability & ops

Built for operators

Every intercept, anomaly score, and ingest request shows up as structured JSON you can ship to Datadog, Prometheus, or whatever you use.

letsping-audit-streamLIVE
10:42:05.331BLOCKEDshell-agentexec_commandrm -rf /home/user/project
10:42:04.892INTERCEPTEDcron-agentheartbeatCost velocity > $0.50/s
10:42:02.147APPROVEDstripe-agentpayout_create$450.00 → acct_x99 (1-tap)
10:41:58.003PATCHEDsocial-botpost_tweetHallucination removed
10:41:55.441ESCROWbilling-bothandoff→ risk-orchestrator (signed)
10:41:52.001INTERCEPTEDfinance-botdb.queryDROP TABLE users;
10:41:48.776TRUST_CHECKdelivery-botattestationScore: 847 · 1,240 fulfilled
10:41:45.230BLOCKEDrag-pipelineupload_fileid_rsa found in payload
_Waiting for next event...
1from letsping import LetsPing
2from langchain.tools import Tool
3 
4lp = LetsPing()
5 
6# Wrap any high-risk tool. LetsPing tracks the full execution graph.
7safe_refund = Tool(
8 name="issue_refund",
9 func=lp.tool("cx-agent", "stripe:refund", "critical"),
10 description="Issues a customer refund via Stripe."
11)
12 
13# Add to your existing graph. Everything else stays untouched.
14graph.add_node("issue_refund", safe_refund)
15 
16# Example intercept:
17# 1. Agent hallucinates a $5K refund.
18# 2. LetsPing catches the off-behavior transition.
19# 3. State parks in Cryo-Sleep. You get a push notification.
20# 4. One tap: approve, reject, or edit the amount.
21# 5. The baseline learns. Your agent gets smarter.

3 LINES · ANY FRAMEWORK

Works with every stack you already use.

FRAMEWORKSTATUS
LangChain / LangGraphNATIVE
Vercel AI SDKNATIVE
CrewAINATIVE
MCP ProtocolNATIVE
OpenTelemetryPASSIVE
Custom AgentsSDK

OpenTelemetry spans are captured passively, so there's no added latency to your LLM calls.

Agent-First · Escrow Contracts

Use signAgentCall, verifyEscrow, and chainHandoff to cryptographically link multi-agent calls. Same firewall + HITL console, with a signed chain-of-custody when agents call each other.

Read the escrow spec →

Your agents' trust layer.
Free to start.

Behavioral firewall. Human approval. Agent credentials.
One SDK, every framework.

No lock-in. One SDK for every agent stack.