How it worksTrustFAQ
AI AGENT GOVERNANCE

The safety layer for AI agents that take actions.

Prevent data leaks and destructive actions by enforcing policy on tool calls — before execution.

ALLOWMODIFYBLOCKESCALATE
Book a demo
At the tool boundary · Shadow mode available · Works with your stack
Audit trace
2026-02-05 14:12:09Z
Tool call
{
  'tool': "db.query",
  'args': {
    "sql": "DROP TABLE customers;"
  }
}
Verdict
BLOCK
DB.DESTRUCTIVE_QUERY

Destructive query stopped before execution.

trace_idtr_9f2c…a13b
policyprod/tool-safety@v12
Tool call → verdict → trace
Audit trace
2026-02-05 14:11:32Z
Tool call
{
  'tool': "email.send",
  'args': {
    "to": "[REDACTED]",
    "body": "Your info..."
  }
}
Verdict
MODIFY
PII_REDACTION

Email address auto-redacted from arguments.

trace_idtr_8a1b…c45d
policyprod/privacy-guard@v4
Tool call → verdict → trace
Audit trace
2026-02-05 14:10:58Z
Tool call
{
  'tool': "stripe.refund",
  'args': {
    "amount": 750.00,
    "reason": "customer_request"
  }
}
Verdict
ESCALATE
HIGH_VALUE_TX

High-value refund routed to human approval.

trace_idtr_3e4f…b21a
policyprod/finance-controls@v2
Tool call → verdict → trace

Get started in 3 commands

git clone → make quickstart → curl
Start here
Running in 5 minutes
Integrate with SDK
Python / TypeScript
Run in shadow mode
Observe without blocking
Turn on enforcement
Shadow → enforce
Policy packs
PII, SQL, payments
See it working
Run the demo

AGENT FRAMEWORKS

LangChain
LangGraph
CrewAI
Google ADK
LlamaIndex
Vercel AI SDK

LLMs

OpenAI
Anthropic
Gemini
Ollama

How it works

NjiraAI sits as middleware between your agent and its tools.

Agent→NjiraAI→Tools / Actions

Intercept tool calls

Traffic flows through the NjiraAI gateway before hitting your APIs.

→ Captures payload + context

Allow / Modify / Block

Policies evaluate arguments against schema, PII rules, and custom logic.

→ Policy evaluation + verdict

Log & Audit

Every verdict is recorded with a full trace for compliance and debugging.

→ Trace + replay + export

Trust & control in production

Built for enterprise teams who need more than prompt engineering.

Shadow Mode
Observe-only, no blocking
Log what would have happened. Validate policies against real traffic before turning on enforcement.
Policy as Code
Versioned, testable, reviewable
Safety rules live alongside your application code. Review in PRs, promote through environments.
Audit Exports
Evidence-ready exports
Full trace logs with verdict, reason code, and payload diffs — structured for SOC 2 / ISO workflows.
RBAC
Scoped access per env / workflow
Granular access controls and scoped API keys per environment, team, or workflow.

Why NjiraAI

Not just prompt filters — tool-call enforcement

Enforce policies on structured tool calls (queries, writes, API requests) at the boundary — not on prompt text.

Not just logs — verdict + auto-correct + replay

Auto-redact PII, auto-sanitize SQL, auto-correct malformed arguments — then log the before/after for replay.

Not just dev-time — production rollout controls

Start in shadow mode to validate without risk. Promote to active enforcement per tool, per environment, when ready.

What teams use NjiraAI for

Security Operations

Data Leakage Prevention

“NjiraAI auto-redacts PII from tool call arguments before the LLM ever sees them, keeping our context window clean.”

Get a risk assessment
Platform Engineering

Preventing Destructive Queries

“We block any SQL query containing DROP/DELETE unless explicitly approved for admin agents.”

Integrate in 30 minutes
AI Product Team

Reliability at Scale

“We use MODIFY policies to auto-correct hallucinated arguments, reducing agent failure rates by 40%.”

Ship with audit trails

Frequently asked questions

You define what's unsafe. NjiraAI ships with default policy packs (destructive queries, PII exposure, high-value transactions), but every rule is customizable. Write your own policies or extend the defaults to match your threat model.

We assume the model can be tricked. NjiraAI enforces policies at the tool-call boundary — so even if injected instructions reach the model, the unsafe tool call is caught and stopped before execution.

Start in Shadow Mode: tool calls flow through normally while NjiraAI logs what it would have done. Once you're confident in your policies, switch to Active mode to enforce verdicts inline. You can switch per tool, per environment.

Yes. NjiraAI can be deployed in your VPC, on-prem, or as a managed service. Reach out to discuss deployment options for strict compliance requirements.

NjiraAI acts as a middleware gateway. Route your agent's tool traffic through our proxy (or use our SDK). Integration typically takes under an hour. Latency overhead is minimized through lightweight decisioning, caching, and async logging.

We log the tool name, input arguments, policy verdict, reason code, and a trace ID. Response bodies are not stored unless you configure it. You can define MODIFY policies that auto-redact fields before logging.

No. Prompt filtering checks text. NjiraAI governs the structured tool calls (JSON function arguments) that agents send to your APIs and databases — preventing unsafe actions, not just unsafe text.

Ready to govern real agent actions?

See NjiraAI intercept, evaluate, and control tool calls in real time.

Book a demo