When AI agents talk to each other in automated pipelines, nobody monitors the conversation. Agent A might say a project costs $1,000. Agent B says $5,000. Neither knows about the contradiction. The wrong number reaches the customer.
Worse: agents fabricate citations that look real. They invent URLs, DOIs, and paper references. They start confident and silently become unsure. One agent's hallucination becomes the next agent's trusted input.
The Solution
InsAIts V2.4 monitors every message between your AI agents and catches problems before they propagate:
5 Hallucination Detection Subsystems:
- Cross-agent fact contradiction tracking (Agent A vs Agent B)
- Phantom citation detection (fake URLs, DOIs, arxiv IDs)
- Source document grounding (verify against your reference docs)
- Confidence decay monitoring (agents losing certainty)
- Self-consistency checking (contradictions within one response)
Plus 6 more anomaly types:
- Shorthand emergence (real words become abbreviations)
- Context loss (topic switches mid-conversation)
- Jargon creation (made-up acronyms)
- Anchor drift (diverging from user's question)
- LLM fingerprint mismatch
- Low confidence detection
Key Features
- Open-source core (Apache 2.0) - anomaly detection, hallucination detection, forensic tracing, dashboard, all integrations
- 3 lines of code to start monitoring
- Privacy-first: All processing runs locally on your machine
- Works with any LLM: GPT-4, Claude, Llama, Gemini, Mistral
- Choose your Ollama model: `insAItsMonitor(ollama_model="phi3")`
- Framework integrations: LangChain, CrewAI, LangGraph
- Ecosystem exports: Slack alerts, Notion, Airtable, webhooks
- Forensic chain tracing: Trace any anomaly to its exact root cause
- Premium features included via pip: Adaptive dictionaries, advanced detection, auto-decipher
- 75+ automated tests covering all detection heuristics
Who Is This For?
- Teams building multi-agent AI systems
- Anyone using LangChain, CrewAI, or LangGraph in production
- Companies where AI accuracy matters (finance, healthcare, legal, e-commerce)
- Developers who want visibility into AI-to-AI communication
I'm the creator of InsAIts. I built this because I kept seeing the same problem across every multi-agent AI system I worked with: agents pass bad information to each other, and there's no monitoring layer to catch it. Today we're open-sourcing the core under Apache 2.0.
The "aha moment" was when I watched a finance pipeline where one agent hallucinated a 5x cost difference. It propagated through three more agents before reaching the output. Nobody caught it because nobody was monitoring the AI-to-AI channel.
InsAIts V2.4 adds deep hallucination detection -- specifically designed for the unique problems that emerge when AI agents communicate:
1. Cross-agent contradictions (the big one -- no other tool catches this)
2. Phantom citations (fabricated URLs, DOIs, paper references)
3. Source grounding (are responses actually based on your documents?)
4. Confidence decay (is the agent losing certainty over time?)
Everything runs locally. We never see your data. The API key is only for usage tracking.
Open-core model: The core (anomaly detection, hallucination detection, forensic tracing, dashboard, all integrations) is Apache 2.0 open-source. Premium features (adaptive dictionaries, advanced detection, auto-decipher) ship with pip install -- proprietary but included in the package. You can also choose your own Ollama model for local processing.
I'd love to hear from anyone building multi-agent systems. What failure modes have you encountered? What would you want monitored?
The Problem
When AI agents talk to each other in automated pipelines, nobody monitors the conversation. Agent A might say a project costs $1,000. Agent B says $5,000. Neither knows about the contradiction. The wrong number reaches the customer.
Worse: agents fabricate citations that look real. They invent URLs, DOIs, and paper references. They start confident and silently become unsure. One agent's hallucination becomes the next agent's trusted input.
InsAIts V2.4 monitors every message between your AI agents and catches problems before they propagate: - Cross-agent fact contradiction tracking (Agent A vs Agent B) - Phantom citation detection (fake URLs, DOIs, arxiv IDs) - Source document grounding (verify against your reference docs) - Confidence decay monitoring (agents losing certainty) - Self-consistency checking (contradictions within one response) - Shorthand emergence (real words become abbreviations) - Context loss (topic switches mid-conversation) - Jargon creation (made-up acronyms) - Anchor drift (diverging from user's question) - LLM fingerprint mismatch - Low confidence detection - Open-source core (Apache 2.0) - anomaly detection, hallucination detection, forensic tracing, dashboard, all integrations - 3 lines of code to start monitoring - Privacy-first: All processing runs locally on your machine - Works with any LLM: GPT-4, Claude, Llama, Gemini, Mistral - Choose your Ollama model: `insAItsMonitor(ollama_model="phi3")` - Framework integrations: LangChain, CrewAI, LangGraph - Ecosystem exports: Slack alerts, Notion, Airtable, webhooks - Forensic chain tracing: Trace any anomaly to its exact root cause - Premium features included via pip: Adaptive dictionaries, advanced detection, auto-decipher - 75+ automated tests covering all detection heuristics - Teams building multi-agent AI systems - Anyone using LangChain, CrewAI, or LangGraph in production - Companies where AI accuracy matters (finance, healthcare, legal, e-commerce) - Developers who want visibility into AI-to-AI communication - Free: 100 messages/day (no API key needed) - Lifetime Starter: EUR99 one-time - 10K messages/day forever - Lifetime Pro: EUR299 one-time - Unlimited foreverFirst 100 users per tier only.
Hi Product Hunt!
I'm the creator of InsAIts. I built this because I kept seeing the same problem across every multi-agent AI system I worked with: agents pass bad information to each other, and there's no monitoring layer to catch it. Today we're open-sourcing the core under Apache 2.0.
The "aha moment" was when I watched a finance pipeline where one agent hallucinated a 5x cost difference. It propagated through three more agents before reaching the output. Nobody caught it because nobody was monitoring the AI-to-AI channel.
InsAIts V2.4 adds deep hallucination detection -- specifically designed for the unique problems that emerge when AI agents communicate:
1. Cross-agent contradictions (the big one -- no other tool catches this) 2. Phantom citations (fabricated URLs, DOIs, paper references) 3. Source grounding (are responses actually based on your documents?) 4. Confidence decay (is the agent losing certainty over time?)
Everything runs locally. We never see your data. The API key is only for usage tracking.
I'd love to hear from anyone building multi-agent systems. What failure modes have you encountered? What would you want monitored?