agentiva 0.1.4

Agentiva — open-source runtime for AI agent safety

Agentiva

Preview deployments for AI agents.

See what your AI agent would do before it does it.

litellm was compromised March 24, 2026 — SSH keys, AWS credentials, and database passwords stolen from 97M monthly downloads. Agentiva catches this class of attack at the action layer.

Quick start (2 minutes)

pip install agentiva
agentiva serve --port 8000
# Open localhost:3000 for the dashboard (from repo: cd dashboard && npm run dev)

Protect your agent (3 lines)

from agentiva import Agentiva

shield = Agentiva(mode="shadow")
tools = shield.protect([your_existing_tools])

# Your agent works exactly the same.
# Every action is intercepted, scored, and logged.

Run the demo

# See 4 real incident recreations
python demo/real_incidents_demo.py

# See PayBot (fintech startup) demo
python demo/paybot_demo.py

# See proof: before vs after comparison
python demo/proof_demo.py

Or use the project venv so dependencies resolve: source .venv/bin/activate then the commands above.

What it catches

Tested against real-world incidents:

• litellm supply chain attack (March 2026) — credential exfiltration blocked
• Amazon Kiro (December 2025) — infrastructure destruction blocked
• Microsoft Copilot (January 2026) — zero-click data theft blocked
• Replit agent (2026) — mass record deletion blocked

Verified results

Benchmark	Result
Agentiva test suite	24,599 tests passing
OWASP LLM Top 10	21/21 (100%)
DeepTeam (Confident AI)	38/47 (80.85%)
Garak (NVIDIA)	2,500 probes scanned
PyRIT (Microsoft)	9/9 scenarios completed

Run benchmarks yourself:

python -m pytest tests/ -m "slow or not slow"  # Full test suite
python benchmarks/run_benchmark.py              # OWASP + incidents
python benchmarks/run_all_benchmarks.py         # All frameworks

Five operating modes

Mode	What it does
Shadow	Observe without executing
Simulation	Preview impact before acting
Approval	Human-in-the-loop for risky actions
Negotiation	Agent learns to self-correct
Rollback	Undo what the agent did

Dashboard

Real-time monitoring at localhost:3000:

• Overview — stats, charts, recent activity
• Live Feed — actions streaming via WebSocket
• Audit Log — searchable history with compliance exports
• Agents — registry with reputation and kill switch
• Policies — YAML rule editor
• Security Co-pilot — ask questions about your agent's behavior

Security co-pilot

Ask naturally:

• "What was blocked?" → real data from your audit log
• "Why was send_email blocked?" → specific tool analysis
• "Is this HIPAA compliant?" → compliance check with regulation citations
• "Is my agent safe for production?" → honest assessment

Basic mode works without any API key. Add OPENROUTER_API_KEY for Claude-powered deep analysis via OpenRouter.

Works with

LangChain, CrewAI, OpenAI Agents SDK, Anthropic, MCP Protocol, or any custom agent via REST API.

Compliance-ready evidence

Generates audit trails aligned with:

• HIPAA — PHI access logs per 45 CFR § 164.312
• SOC2 — Evidence for CC6-CC8 controls
• PCI-DSS — Cardholder data monitoring per Req 3, 7, 10

Note: Agentiva helps prepare for compliance audits. Certification requires a third-party assessor.

Pricing

Tier	Price	Agents
Free	$0/forever	1 agent
Pro	$18/month	Up to 3
Team	$54/month	Unlimited
Enterprise	Custom	Custom

Self-hosted is free forever. Cloud dashboard on waitlist.

Architecture

┌────────────────────┐
│  AI Agent          │  LangChain · CrewAI · OpenAI · MCP · custom tools
└─────────┬──────────┘
          │ tool_call
          ▼
┌─────────────────────────────────────────────────────────────────────┐
│                        AGENTIVA API (FastAPI)                        │
│  /api/v1/intercept · /api/v1/audit · /api/v1/chat · WebSocket feed   │
└───────────────────────────────┬─────────────────────────────────────┘
                                │
          ┌─────────────────────┴─────────────────────┐
          ▼                                           ▼
┌──────────────────────┐                 ┌──────────────────────┐
│  Interceptor         │                 │  Shield Chat          │
│  PolicyEngine (YAML) │                 │  Sessions + messages  │
│  SmartRiskScorer     │                 │  (SQLite persistence) │
│  PHI detector        │                 │  + optional LLM layer │
│  Behavior / drift    │                 └──────────┬────────────┘
└──────────┬───────────┘                            │
           │                                        │
           ▼                                        ▼
┌──────────────────────┐                 ┌──────────────────────┐
│  Modes               │                 │  Compliance KB        │
│  Shadow · Approve ·  │                 │  HIPAA · SOC 2 ·      │
│  Live · Negotiation  │                 │  PCI-DSS citations +  │
│  Simulator · Rollback│                 │  evidence SQL hooks   │
└──────────┬───────────┘                 └──────────────────────┘
           │
           ▼
┌─────────────────────────────────────────────────────────────────────┐
│  Persistence: action_logs (audit) · agent registry · approvals ·     │
│  chat_sessions / chat_messages                                       │
└─────────────────────────────────────────────────────────────────────┘
           │
           ▼
┌──────────────────────┐
│  Tools / APIs        │  Email · DB · Slack · shell · payments…
└──────────────────────┘

API reference (short)

Endpoint	Method	Description
/health	GET	Health check + mode + risk threshold
/api/v1/intercept	POST	Intercept an agent action
/api/v1/audit	GET	Query audit log
/api/v1/report	GET	Summary report
/api/v1/settings	PUT	Runtime mode + risk threshold
/ws/actions	WebSocket	Real-time action stream

Full OpenAPI at http://localhost:8000/docs.

Testing

python -m pytest tests/ -q
python -m pytest tests/ -m "slow or not slow" -q

Contributing

See CONTRIBUTING.md.

License

Apache 2.0 — see LICENSE.

Built by

Rishav Aryan — ML Engineer, George Mason University

GitHub · Twitter