Let agents act. Stay in control of every action.
AgentTrust sits between your AI agents and the tools they call. Every agent gets a verifiable identity, and every action is checked against your policies and allowed, denied, or flagged before it runs.
Works with any framework — LangChain, Vercel AI SDK, CrewAI, or your own loop.
How it works
One HTTP call stands between an agent's intention and a real-world action. That is the whole trust layer.
1. Agent wants to act
Before doing anything consequential — sending money, writing to a database, emailing a customer — your agent pauses and asks AgentTrust for permission.
2. Identity is verified
The agent sends its API key. AgentTrust confirms which agent is calling, that it is active, and attributes the request to it.
3. Policies are evaluated
The action, resource, and parameters are matched against your rules. Conditions like amount > 1000 decide whether to allow, deny, or flag for review.
4. Decision is logged & returned
The verdict is written to an immutable audit log and returned in milliseconds. Your agent proceeds only if the answer is allow.
One call, one clear answer
Your agent asks before it acts. AgentTrust replies with a decision and the reason behind it.
Agent asks
POST https://www.agenttrustcore.com/api/evaluate
Authorization: Bearer atk_live_your_agent_key
{
"action": "payment.send",
"resource": "stripe",
"params": { "amount": 5000, "currency": "usd" }
}AgentTrust decides
{
"decision": "deny",
"reason": "Payments over $1,000 require human approval",
"policy": "High-value payment guard",
"agent": "Support Agent",
"logId": "log_a1b2c3"
}Everything you need to trust an agent
Identity and guardrails that are simple to reason about and hard to bypass.
Agent identity
Register every agent and issue it a scoped API key. Suspend a misbehaving agent instantly and all of its actions stop.
Policy guardrails
Match on action, resource, and conditions like amount or recipient. Choose allow, deny, or flag for human review.
Immutable audit log
Every request and decision is recorded with its reason, so you can prove exactly what each agent did and why.
Provider-agnostic
A plain HTTP API plus a zero-dependency SDK. Drop it into any agent framework or a custom loop in minutes.
Fail-closed by default
Unknown agents, expired keys, and unmatched actions are denied unless you explicitly allow them.
Portable storage
Runs in-memory for testing today and swaps to a managed database later without changing a single API call.
Built for high-stakes actions
Wherever an agent can cause real damage, put a decision in front of it.
Payments & refunds
Cap the amount an agent can move, require approval above a threshold, and block unknown payees.
Data access
Allow reads but gate deletes and schema changes, scoped to specific tables or resources.
Outbound messages
Let agents draft freely but flag or deny sends to customers outside an approved domain list.
Tool & command execution
Whitelist safe tools and deny destructive commands before they ever reach your systems.
Give your agents an identity and a rulebook
Register your first agent, write a policy, and watch decisions land in the audit log — in a couple of minutes.