See every AI agent. Govern every action.

The Complete Agent Control Plane

Four integrated pillars that give you full visibility, intelligent detection, cost control, and automated governance across your entire AI agent ecosystem.

Pillar 1

Agent Discovery & Registry

Every agent that emits OpenTelemetry to MeshAI enters the registry automatically, whether or not anyone registered it. Point your OTel Collector at MeshAI once, and each agent any team deploys inside that pipeline shows up on its own, with framework, model, cost, and behavior attached. The result is a living, searchable registry with no per-agent onboarding.

  • Zero-Touch Registration

    Agents enter the registry the first time they emit telemetry. No manual onboarding, no per-agent setup.

  • Unregistered Agent Flagging

    Agents that appear in telemetry but were never declared are tagged for review, so sprawl inside your pipeline surfaces on its own.

  • Health Dashboard

    Real-time status, uptime, and performance metrics for every registered agent

  • Framework-Agnostic Catalog

    Unified metadata regardless of whether agents use LangChain, CrewAI, AutoGen, or custom frameworks

  • Telemetry-Scoped by Design

    Discovery covers everything in your OpenTelemetry pipeline. MeshAI installs no endpoint agents and scans no networks.

customer-support-agentHealthy
CrewAI | OpenAI GPT-4 | 2.3K req/day
data-analysis-pipelineDegraded
LangGraph | Anthropic Claude | 890 req/day
unknown-slack-botUnregistered
Framework undeclared | OpenAI | 450 req/day | Auto-discovered

Live Detection Dashboard

Cost Spike1 alert
Z-score > 3.0 or 5x mean
Reliability DecayClear
Error rate 2x baseline
Behavioral Drift3 alerts
New model: claude-3-opus
Security Threat2 alerts
10x request rate spike
Runs every 5 minutes | Webhook alerts | Configurable thresholds
Pillar 2

ML-Powered Anomaly Detection

Four detection algorithms analyze every agent's behavior, costs, reliability, and security patterns every 5 minutes. Configurable alert rules with webhook delivery for Slack, PagerDuty, and custom endpoints.

  • Cost Spike Detection

    Z-score analysis against 24-hour rolling baseline detects unusual token spend. Absolute multiplier catches sudden 5x+ cost jumps.

  • Reliability Decay

    Monitors error rates and P95 latency against 7-day baselines. Alerts when error rate exceeds 2x normal or 25% absolute.

  • Behavioral Drift

    Detects when agents start using models not seen in the past 7 days, a strong indicator of configuration changes or compromise.

  • Security Threats

    Catches request rate spikes (10x+ normal) and dormant agent reactivation: agents silent for 7+ days suddenly making requests.

Pillar 3

Cost Intelligence & Optimization

Know exactly where every token dollar goes. Attribute costs to teams, projects, and individual agents. Set guardrails and let ML optimize your model routing.

  • Token-Level Attribution

    Granular spend tracking by team, project, agent, and individual request

  • Budget Guardrails

    Automatic enforcement with configurable thresholds and escalation policies

  • Model Routing Optimization

    Route requests to the most cost-effective model that meets quality requirements

  • ML Spend Forecasting

    Predict future costs based on usage trends and planned agent deployments

Monthly Cost Breakdown

Engineering$4,280 (42%)
Customer Support$2,890 (28%)
Marketing$1,540 (15%)
Sales$1,020 (10%)
Unattributed$510 (5%)
Total$10,240/mo

Policy Engine: Live

POST /api/v1/policies
{
"name": "Production models only",
"policy_type": "model_allowlist",
"conditions": { "environments": ["production"] },
"rules": { "allowed_models": ["gpt-4o", "claude-3-sonnet"] }
}
EU AI Act - High Risk ClassificationCompliant
HITL Approval Required2 Pending
Pillar 4

Governance & Compliance Engine

Eight policy types enforced in real-time through the proxy with sub-5ms overhead. Immutable audit trails, EU AI Act readiness scoring (0-100), and HITL approval workflows, all built and deployed.

  • Real-Time Policy Enforcement

    8 policy types (model allowlist, block provider, require approval, budget limit, rate limit, human review, prompt injection, PII filter) evaluated at the proxy layer

  • Prompt Injection Protection

    15+ injection patterns scanned on every request. Role override, jailbreak, delimiter attacks, and encoding evasion, blocked before reaching the LLM.

  • PII Detection & Redaction

    8 PII types detected in LLM responses (email, phone, SSN, credit card, IP, passport, IBAN). Block, redact, or allow with logging: your choice.

  • EU AI Act Readiness Score

    Automated 0-100 compliance score across 5 components: audit trail, risk classification, HITL, documentation, data retention

  • Human-in-the-Loop Approvals

    Proxy returns 403 for approval-required policies. Dashboard queue with approve/deny. Redis-cached for instant subsequent access.

  • Risk Classification

    AI-assisted risk suggestion from agent metadata with mandatory human confirmation per Article 14. FRIA templates and transparency cards.

Detection Architecture

Statistical detection runs in real-time on lightweight Cloud Run workers. BigQuery ML powers daily ARIMA cost forecasting. OpenTelemetry ingestion connects any agent framework automatically.

Real-Time Statistical Detection

Z-score, rate-of-change, and set-diff algorithms run every 5 minutes on 5-minute metric aggregations. Sub-$30/mo infrastructure cost.

BigQuery ML Forecasting

ARIMA_PLUS time-series models trained weekly per agent detect cost anomalies against predicted spend. ML.DETECT_ANOMALIES runs daily.

OpenTelemetry Ingestion

OTLP/HTTP endpoint auto-discovers OpenClaw and NemoClaw agents from trace data. Zero code changes: one env var to connect.

Ready to Take Control?

Sign up for a free account and deploy the Agent Control Plane today.

See these pillars applied to real-world use cases, read what AI agent observability covers on its own, or see how the runtime layer sits against device, observability, and GRC tooling in the layer-by-layer comparisons.