Plurai
by Plurai
Simulation testing for AI agents: digital twins, thousands of synthetic users and cheap small-model guardrails
Plurai is an AI agent evaluation platform for engineering teams shipping agents to production. It builds a sandboxed digital twin of an agent's tools and data, runs thousands of simulated users against it to find hallucinations, policy breaches and broken tool calls, then trains small models that act as low-cost production guardrails.
Plurai was founded by CEO Dr. Ilan Kadar and CTO Dr. Elad Levi, who previously built AI systems at the automotive technology company Nexar. Its product has two parts. Plurai Simulation maps an agent as built, mocks the tools, sub-agents and knowledge sources around it in an isolated sandbox, and runs multi-turn scenarios, including adversarial ones, to surface hallucinations, policy violations, broken tool calls, failed tasks and sensitive-data leaks; each failure comes back with a proposed fix that is re-verified in the same environment. It installs as a plugin inside Claude Code, GitHub Copilot CLI or Codex CLI (Cursor is listed as coming soon), with a Python SDK for CI and support for LangGraph, CrewAI, OpenAI Agents SDK, Google ADK, LlamaIndex, Mastra and Salesforce Agentforce, and it exports traces to LangSmith, Langfuse, Arize and Braintrust. The second part, post-training, uses the simulation data to train small language models served on NVIDIA Nemotron and NIM as evaluators and guardrails with under 100 ms latency. Plurai also maintains IntellAgent, an Apache-2.0 open-source agent simulation framework with about 1,300 GitHub stars. The company raised a $13 million seed led by Team8, reported by Ynet in November 2026, and launched simulation to all developers with a free tier. Gartner listed it as a representative vendor in its 2026 Market Guide for AI Evaluation and Observability Platforms, and its BARRED synthetic-data method was accepted at ICML 2026.
An AI engineering lead about to put a customer-facing or tool-using agent into production who needs evidence it won't break policy or misuse tools.
Agent failures are found in a mocked sandbox before customers hit them, and production guardrails run on small models costing a fraction of a frontier-model judge.
At a Glance
- Category
- Developer Tools
- Pricing
- Freemium, Subscription, Usage-based
- Target Market
- AI Engineers, CTOs, Enterprise Developers, Product Managers
- Deployment
- Cloud-first, API-based
Key Features
- ✓Agent digital twin
Mocks and seeds every tool, sub-agent and knowledge source the agent calls, so tests run without touching real systems or writing real data.
- ✓Synthetic user simulation
Runs thousands of multi-turn and adversarial scenarios to find hallucinations, policy violations, broken tool calls and data leaks before launch.
- ✓Verified fixes
Each failure returns with a proposed fix that is re-run in the same sandbox, so teams know the change works before shipping it.
- ✓Coding-agent plugins
Installs into Claude Code, GitHub Copilot CLI and Codex CLI with a single simulate command, keeping testing inside the developer's workflow.
- ✓Small-model guardrails
Trains custom small language models from simulation data that serve as evaluators and guardrails at under 100 ms latency and low token cost.
- ✓Observability integrations
Analyses and replays traces from LangSmith, Langfuse, Arize and Braintrust and records runs as experiments there, fitting existing eval stacks.
Capabilities
Use Cases
- •Pre-production agent QA
A team points Plurai at its support agent, runs simulated customers through refund and escalation flows, and fixes policy breaks before go-live.
- •Regression testing after changes
After a prompt or model change, teams rerun the saved scenario suite against the existing digital twin to catch regressions quickly.
- •Cheaper production guardrails
A company replaces a frontier-model judge with a Plurai small model, cutting per-call evaluation cost while keeping latency under 100 ms.
- •Grounding tests in real traffic
Production traces from an observability tool seed the scenario suite, so simulated users reflect how real customers actually talk to the agent.
Ideal For
Best For
- ✓Pre-launch testing of customer-service and tool-calling agents against thousands of synthetic multi-turn conversations
- ✓Teams already building in Claude Code, Copilot CLI or Codex who want simulation from inside the coding agent
- ✓Replacing expensive LLM-as-judge evaluators with fine-tuned small models for always-on guardrails
- ✓Regulated companies that must show policy compliance testing before deploying an agent
Not Ideal For
- ✗Teams that need CI pipeline integration today, since headless CI runs are listed as coming soon
- ✗Cursor-first shops, because the Cursor plugin is not yet available
- ✗Buyers who require on-premise deployment without an enterprise contract; on-prem is gated to the Business/Enterprise tier
Integrations
Deployment
Market Analysis
Pros
- ✓Free tier with no overage charges and published per-plan pricing
- ✓Broad framework coverage, from LangGraph and CrewAI to Agentforce and n8n
- ✓Open-source IntellAgent framework (Apache-2.0, about 1.3k GitHub stars) shows the underlying method
- ✓Gartner representative vendor for AI evaluation and observability in 2026
Cons
- ✗Product Hunt's only review says the synthetic test-case review UI is slow and the docs are thin in places
- ✗CI pipeline runs and the Cursor plugin are still listed as coming soon
- ✗Headline cost and quality claims (for example 44x cheaper than Claude Haiku 5.5) come from vendor benchmarks on a single workload
- ✗Compliance detail is vague: the pricing page cites AICPA verification without naming the SOC report type
Pricing
Simulation Free
$0
- ✓12,300 one-time credits
- ✓2 agent setups
- ✓3 one-time runs
- ✓10 sandbox hours
- ✓No overage charges
Simulation Base
From $60/mo
- ✓28,710 credits/month
- ✓20 runs/month
- ✓30 sandbox hours/month
- ✓Pay-as-you-go at $0.02/credit
Simulation Business
From $210/mo
- ✓65,570 credits/month
- ✓60 runs/month
- ✓100 sandbox hours/month
- ✓Pay-as-you-go at $0.015/credit
Plurai SLM (pay as you go)
$0.0078 per 1M tokens
- ✓Under 100 ms latency
- ✓Up to 20 endpoints
- ✓Unlimited seats
Business / Enterprise
Contact for pricing
- ✓On-prem deployment
- ✓Enterprise SSO
- ✓Custom inference pricing and SLA
Plurai publishes list pricing (checked November 12, 2026). Simulation is metered in credits, where one credit is one cent of model-provider usage at list price, plus runs and sandbox hours; Base is $60/month and Business $210/month, with admin-capped overage and unlimited members. Guardrail models are billed separately per token ($0.0078 per 1M for the SLM, $0.3 per 1M for the optimized LLM). On-prem, SSO and SLAs require a custom Enterprise contract.
Security & Compliance
Connect
Sources
This page was written from 7 sources, 5 on domains other than plurai.ai.
Stay Ahead of the Curve
Weekly enterprise AI insights for technology leaders. No spam, no vendor pitches—unsubscribe anytime.
SubscribeRelated Products
Context.dev
One web context API for AI agents: scrape, crawl, search, extract and enrich with a single key
TestGrid ONE
AI testing agents plus real-device and browser infrastructure, deployable on-premises or air-gapped
Ampersand
Integration infrastructure that lets AI agents read and write in customer CRMs and ERPs
OutSystems Agent Experience
Governed AI development on OutSystems for Claude Code, Cursor, Codex and Kiro, via MCP