insa-its 4.8.7

Open-core multi-LLM communication monitoring, hallucination detection & deciphering for agent systems

InsAIts — Runtime Security for Multi-Agent AI

See what your agents say to each other. Stop the bad stuff before it ships.

10,000+ installs from PyPI. Integrated into the AgentShield runtime — a community fork of Anthropic's Claude Code ecosystem with 35k+ stars.

Live threat dashboard — session score, anomaly feed, agent intelligence, OWASP alignment. Visit the website →

Latest demo • Dashboard walkthrough • Prototype

---

The problem

AI-to-AI communication is a blind spot. Microsoft, CrowdStrike and Cisco can see that an agent exists — they cannot see inside a spawn chain. When an agent spawns a subagent that spawns another subagent, the behavior of those deeper agents is invisible to every enterprise monitoring tool on the market today.

Meanwhile, agent-to-agent interactions fail silently. Hallucinations propagate as "facts." Agents invent citations. Sensitive data leaks into tool chains. Meaning drifts across messages. Tool descriptions change between discovery and invocation. The agents nobody watches are the agents that cause the most damage.

InsAIts makes this visible and acts on it.

---

What it does

A runtime security layer that plugs into your agent stack. It observes every message and every tool call, detects a wide class of multi-agent failures and attacks, and can actively intervene — blocking, quarantining, rerouting, or escalating to a human operator.

• Full spawn-tree visibility — every agent, every subagent, every tool call. Attribution down to the exact caller. No blind spots.
• Broad runtime coverage — aligned with the OWASP MCP Top 10 and Agentic AI Top 10. Catches credential exposure, prompt injection, tool manipulation, data exfiltration, rogue behaviour, hallucination chains, covert channels, and more.
• Active protection, not just alerting — critical anomalies can be quarantined, rerouted, or escalated to a human. Configurable per severity.
• 100% local by default — your agent traffic never leaves the box. No cloud round-trips. Audit logs store hashes, not raw content. GDPR-ready.
• Tamper-evident audit trail — hash-chained so any modification is detectable. Export-ready for compliance review.
• Live dashboard — real-time threat score, agent intelligence, anomaly stream, circuit-breaker grid, OWASP coverage.
• Drop-in integrations — LangChain, CrewAI, LangGraph, AutoGen/ag2, and direct Claude Code / Cursor / Windsurf / Codex / Aider hook support.

---

Install

pip install insa-its[full]

Quick start

# Terminal 1 — event hub
insaits-collector

# Terminal 2 — live dashboard
insaits-dashboard

# Open http://localhost:5001

Three lines to integrate in Python:

from insa_its import insAItsMonitor

monitor = insAItsMonitor()
result = monitor.send_message(
    text=agent_response,
    sender_id="OrderBot",
    receiver_id="InventoryBot",
    llm_id="gpt-4o",
)
# Act on the result — see the user guide for the full API.

That's the whole public surface you need to start. The rest of the behaviour is configured through the dashboard or .insaits_config.json — no code changes required.

Claude Code integration

One hook entry in .claude/settings.json gives you pre-execution gating, full tool-call audit, spawn-tree attribution, and session continuity across compaction. See the user guide for the exact snippet.

Session-length multiplier. InsAIts captures a lightweight session vault after each tool call. After context compaction, the next session resumes from a sub-kilobyte brief instead of re-reading thousands of tokens of history. Users report Claude Code Opus sessions going from ~40 min to 3h+ on the same Pro plan.

Minimal install (no local embeddings):

pip install insa-its

---

Pricing

All detection features are available under open-core (Apache 2.0). You pay for productivity, longer sessions, premium detectors, and team features.

Tier	Monthly	Lifetime	What you get
Trial	free	14 days	Full feature access. No card required.
Starter	€10	€99	Full detection + reliability gates + session vault + dashboard.
Pro	€49	€299	Starter + advanced context engineering, inter-session dialog, forensic export, Decipher engine, priority support.
Enterprise	from €200	custom	SOC2-ready audit export, multi-seat, white-label dashboard, dedicated support.

Activate

Pick a link. Your license key arrives by email after checkout.

• Starter — €10/month
• Pro — €49/month
• Starter Lifetime — €99
• Pro Lifetime — €299

Install:

export INSAITS_LICENSE_KEY="insaits_live_..."
insaits-collector

Enterprise and custom: [e-mail].

Passive mode

When the trial lapses and no key is set, detection still runs and anomalies still surface on the dashboard. What unlocks with a paid key is active intervention — the corrective signals the layer sends back to your agents. You always see everything InsAIts catches.

---

Coverage

InsAIts covers both the OWASP MCP Top 10 and the OWASP Agentic AI Top 10. Threat types are grouped into:

• Security — credential leakage, prompt injection, tool poisoning, data exfiltration, unauthorised access, covert channels, shadow servers.
• Hallucination & grounding — cross-agent contradictions, phantom citations, ungrounded claims, confidence decay.
• Semantic & communication — meaning drift, emergent shorthand, context loss, inter-agent jargon.
• Behavioural — rogue deviation from baseline, probing patterns, rapid tool-call anomalies, loop detection.
• Reliability — premature completion claims, unverified assertions, file-reference hallucinations, escalation on repeated failures.

Detection is implemented across many adapters and detectors — the exact set is under open-core and can be toggled per deployment. The premium tier adds semantic divergence scoring, advanced rogue-subagent tracking, and the compliance export engine.

---

Live dashboard

The web dashboard renders in real time on http://localhost:5001:

• Session threat score with stage label (safe → watch → alert → isolate)
• Live anomaly stream with severity colour coding
• Per-agent intelligence and trust scores
• Circuit-breaker grid for blocked agents
• OWASP alignment panel
• Tool-call inspector with full spawn-tree drill-down
• Token-usage panel with cost estimates

A companion terminal dashboard is available for VS Code split-panes:

pip install insa-its[dashboard]
insaits-tui

---

Audit & compliance

• Hash-chained audit log — SHA-256 per entry, verifiable integrity.
• Prometheus metrics endpoint for existing monitoring stacks.
• Forensic export of rogue agent behaviour — timeline, intent, threat score.
• SOC2-ready evidence chain (Pro / Enterprise).
• 100% local processing. Audit logs store hashes, never raw content. API keys are hashed before storage. GDPR-aligned by default.

---

Integrations

LangChain • CrewAI • LangGraph • AutoGen / ag2 • Claude Code • Cursor • Windsurf • Codex CLI • Aider • Continue.dev • Kiro • Slack notifier • Notion export • Airtable export • Prometheus metrics.

Wire-up for each integration is documented in the user guide. A typical integration is a one-liner wrap on your chain / crew / graph object.

---

Architecture (high-level)

Your multi-agent system
        │
        ▼
┌─────────────────────────┐
│   InsAIts runtime layer │    ← detection, intervention, circuit-breakers
└─────────────────────────┘
        │
        ├── local event hub (port 5003)
        ├── live dashboard  (port 5001)
        └── tamper-evident audit log (local disk)

Everything runs on your machine. No message content leaves the box.

---

Use cases

Industry	What InsAIts catches
E-commerce	Order bots losing context mid-transaction
Customer service	Agents developing incomprehensible shorthand
Finance	Analysis pipelines hallucinating metrics between agents
Healthcare	Multi-agent decisions where errors have consequences
Research	Fabricated citations, phantom papers
Legal	AI-generated documents with ungrounded references
Security	Red-team AI exercises, exfiltration attempts, compliance
DevOps	Claude Code / Copilot tool calls in CI/CD pipelines

---

What's new

Each release focuses on a specific reliability or cost axis. The most recent work targeted token efficiency under long multi-agent runs — corrective signals are now more surgical, which reduces Claude-context overhead during continuous monitoring without sacrificing coverage. Full notes on the GitHub releases page.

---

Documentation & support

• User guide + installation: installation_guide.md
• Live API reference: insaits-api.onrender.com/docs
• Privacy policy: PRIVACY_POLICY.md
• Terms of service: TERMS_OF_SERVICE.md
• Email: [e-mail]
• Issues: GitHub Issues
• API status: insaits-api.onrender.com

---

License

Open-core model.

• Core SDK (Apache License 2.0) — free to use, modify and distribute.
• Premium features (shipped with pip install insa-its) — proprietary, gated by license key.

---

InsAIts — Making multi-agent AI trustworthy, auditable and secure.
OWASP MCP + Agentic AI Top 10 coverage • Full spawn-tree visibility • Tamper-evident audit • 100% local • 1900+ tests passing.

Starter from €10/month · Pro from €49/month · Enterprise custom
Contact: [e-mail]