Trust your AI agents in production.
Agents are easy to demo and hard to trust — the moment one can move money, change access, or delete data, a single wrong call is an incident. Kelhe scores every action your agents take before it runs and returns a decision in milliseconds: allow, hold for a human, or block.
Wrap your tools once · Start in observe mode · No agent rewrites
assign_roles(user="Suraj", roles=["admin"])
The blocker to agents in production isn't capability. It's trust.
Nearly every company is piloting agents. Almost none trust them enough to let them act in production — because nothing independent decides what an agent is allowed to do once it's loose on real systems.
of enterprises are piloting AI agents.
trust them enough to run in production. Closing that gap is the whole job.
barrier to scaling agents is now security & risk — ahead of cost or capability.
Sources: 2026 enterprise surveys on agentic AI adoption and the OWASP Top 10 for Agentic Applications (“excessive agency”).
One independent control point on every action an agent takes.
Not a smarter prompt, and not a dashboard you read after the fact. Kelhe sits on the path of the action itself — outside the model — and decides, in real time, whether it runs.
See every action
Every tool call is scored and recorded — a complete, queryable answer to “what are our agents actually doing,” from the first day you install it.
Score the risk
The engine reads the real action — the tool, its arguments, the agent's mandate and intent — and returns one risk score, 0 (safe) to 10 (dangerous).
Decide before it runs
Above your threshold, the action is held for a human or blocked — never after the fact. Every outcome lands in the record with a reason.
Your agent calls a tool. Kelhe decides if it runs.
The wrapper captures the real action and its context, the engine scores it, and the tool executes only on a decision — so enforcement happens outside the model, where it can't be talked out of it.
Wrap one tool. Start in observe mode. Enforce when you trust it.
There's no platform to migrate to and no agent to rewrite. You wrap the tools your agents already call — one line each — and Kelhe begins scoring. This is the path security teams recommend: watch first on your highest-risk workflow, turn on enforcement when the decisions earn it.
# pip install "kelhe[openai]"
from kelhe import Kelhe
from kelhe.adapters.openai_agents import guard
from agents import Agent
kelhe = Kelhe(api_key="sk_kelhe_…") # observe-only by default
def issue_refund(amount: int, customer: str) -> str:
"Issue a refund to a customer."
return f"refunded ${amount} to {customer}"
agent = Agent(
name="billing",
instructions="process approved customer refunds",
tools=[guard(kelhe, issue_refund)], # ← the only change
)
Same one-line guard() across
OpenAI Agents SDK, LangGraph / LangChain,
CrewAI, and AutoGen.
Drop it in. Change nothing.
In observe mode Kelhe scores and logs every action but blocks nothing — so adding it to a live system is risk-free. You get a complete record of what your agents do and proof of which actions would have been stopped.
Flip one flag. Now it acts.
Set mode=ENFORCE and a threshold. Risky actions are held for a
human or blocked before they run, routed into the approval and on-call tools
you already use — so a person owns every irreversible call.
The actions that turn a wrong call into an incident.
Agents fail in ways a prompt can't prevent — what the OWASP LLM Top 10 calls excessive agency. Kelhe evaluates the actual call, so the high-consequence actions are the ones that get scored, held, or blocked.
Privilege & access changes
Role grants, permission escalations, and IAM edits — held for human approval before anyone gains access they shouldn't have.
Money movement
Payments, transfers, refunds, and payouts scored against limits and context before any funds leave.
Production & infrastructure
Deploys, configuration changes, and infrastructure edits checked before they touch a live environment.
Bulk & sensitive data access
Large or sensitive reads and exports flagged before data can be exfiltrated at scale.
Destructive operations
Deletes, drops, and overwrites stopped before an irreversible change lands.
Customer-facing actions
Outbound emails, messages, and account changes reviewed before they reach a real customer.
One control point, for everyone accountable for the agent.
From the solo founder shipping their first agent to the buying committee at a Fortune 20 — the same wrapper, read by each team for what they own.
Builders & startups
Ship the agent without hand-rolling guardrails. One line per tool, observe mode free to start — safety you can add the same afternoon.
Security & platform
An independent enforcement point in front of every agent action, shared across frameworks — defense that doesn't depend on the model behaving.
Risk & compliance
Demonstrable human oversight and an exportable record of every decision and reason — evidence you can put in front of auditors.
Built to be trusted on the path of your most sensitive actions.
Kelhe sits in front of the actions that matter, so it's engineered for the controls a platform, security, and procurement team expects.
Encrypted in transit
All traffic to the scoring API is encrypted with TLS. You control what context is sent and redact sensitive data at the wrapper, before it leaves your environment.
Never used to train
Your tool calls and context are used only to return a decision — never to train models. Data minimization is the default.
Configurable fail-safe
Choose fail-open or fail-closed per environment, so availability always matches your risk tolerance — your decision, not ours.
Audit & SIEM export
Every score, decision, and reason is recorded, queryable, and exportable to Splunk or your SIEM — a defensible record for every action.
SSO & least-privilege access
SSO/SAML and role-based policies per team, environment, and tool. Internal access to customer data is restricted on a least-privilege basis.
SLAs & data residency
Uptime and latency SLAs, region pinning, and private connectivity options for enterprise agreements — and a DPA on request.
- SOC 2 Type II — in progress
- GDPR-aligned
- Data-minimization by design
- DPA & security review — on request
- NIST AI RMF
- OWASP LLM Top 10 — Excessive Agency
- EU AI Act — human oversight
- ISO/IEC 42001
Kelhe is designed to support the controls these frameworks ask for — it maps to them; it is not a certification. See how it maps.
Simple usage-based pricing
Pay only for evaluation volume. No tiers, no feature gating.
Metering is based on the evaluated tool payload and the generated risk audit. All operating modes are included without tiering.
Request accessVolume & enterprise agreements — talk to sales.
What review teams ask first
How much latency does Kelhe add?
Kelhe is built for the hot path. Scoring runs as a single hosted API call designed to return a decision in milliseconds, so guarded tool calls proceed with minimal added latency. Enterprise agreements include latency SLAs.
What happens if Kelhe is unreachable?
Failure behavior is configurable per environment. Choose fail-open — proceed if scoring is unavailable — or fail-closed — block until a decision is returned. Availability always matches your risk tolerance.
Do you train models on our data?
No. Customer tool calls and context are used only to return a decision, never to train models. Kelhe follows data-minimization principles, and a DPA and security review are available on request.
Can Kelhe run in our VPC or a specific region?
Deployment and data-residency options, including region pinning and private connectivity, are available for enterprise agreements. Email team@kelhe.com to discuss requirements.
Why not just handle this in the prompt?
Prompts shape intent; they don't enforce it. A model can be persuaded, confused, or wrong, and a prompt leaves no independent record of what was allowed. Kelhe evaluates the actual tool call at runtime — outside the model — and produces a decision and an audit trail you can rely on.
How does Kelhe map to our AI governance frameworks?
Kelhe is designed to support the controls these frameworks ask for: independent runtime enforcement and measurement of agent actions (NIST AI RMF), mitigation of excessive agency (OWASP LLM Top 10), demonstrable human oversight of consequential actions (EU AI Act), and an auditable management record (ISO/IEC 42001). Kelhe maps to these standards; it is not itself a certification.
Can we run a security review before we commit?
Yes. We share a security overview, complete security questionnaires, and provide a DPA on request. Private beta teams are onboarded individually, so a security and procurement review is part of how we start — email team@kelhe.com.
Put Kelhe in front of your agents.
Tell us what your agents are doing and we'll set you up. Private beta — we onboard teams individually so the fit is right.
-
1
Reach out
Send the short form, or email us directly. We reply within one business day.
-
2
30-minute scoping call
We map your agents, tools, and the actions that carry real risk.
-
3
Guided pilot
Start in observe mode on your own workloads, then flip to enforce when you're ready.