Agentic AI securityAn MCP gateway for every agent call

Don’t just govern your agents. Stop them mid-call.

Every agent→MCP call passes through the AtmanID gateway inline — checked against the agent’s scope and blocked before it reaches the server, with humans and agents on one tamper-evident audit trail.

Click or tap an icon to explore.

Checked inline, on every call

  • token
  • agent status
  • owner
  • scope
  • rate limit
  • data-loss rules
  • approval holds
AtmanID MCP gateway · live decision
  1. agentInvoice Triage Agent → netsuite.update_vendor_bank_details
  2. gatewaytoken valid · agent ACTIVE · owner: Finance
  3. gatewayscope check ✗ tool is outside the agent's granted scope
  4. resultDENY · INSUFFICIENT_SCOPE · never reached the NetSuite server
  5. auditdecision recorded on the hash-chained audit trail
An agent with a legitimate token asks for something it was never scoped to do. The gateway refuses inline, before the call reaches the server, and the decision is on the audit trail.MCP is how AI agents call tools such as GitHub, Slack and NetSuite. What is an MCP gateway?

AtmanID in a minute.

Why agents need identity and access management, and how AtmanID checks every call they make. See it in the product →

Read the transcript
  1. Your AI agents already have access. To your code, your customers, and your money. And nobody approved most of it.
  2. AtmanID checks every call an agent makes, before it reaches the tool.
  3. An agent tries to change a vendor’s bank details. Refused. The same agent reads an invoice. Allowed. Decided per call, not per agent.
  4. People and agents share one identity model. One lifecycle. One set of access reviews. And you can ask who can do what, and why.
  5. Every allow, every deny, every approval goes on one hash-chained record. When the auditor asks, evidence is a query.
  6. AtmanID. Identity and access management for people and AI agents.

Compliance posture — stated honestly, verifiable on the Trust page

  • SOC 2

    Not yet audited

    Self-assessed; detail on the Trust page

  • NIST 800-53

    Controls mapped

    Per-control coverage matrix

  • GDPR

    Tooling built in

    Erasure · retention · DPA on request

  • Tenant isolation

    Database-enforced

    Row-level security, probed per schema change

  • Audit chain

    Tamper-evident

    Hash-chained, append-only

  • API contract

    OpenAPI 3.1

    Semver · 6-mo deprecation

Capabilities

Built for the moment an agent acts.

Identity tools decide who gets a token. AtmanID also decides what happens next: every call, every lifecycle change, on one record.

Inline enforcement

Stop the call, not just the credential.

Every agent-to-tool call passes through the AtmanID gateway and is checked against that agent’s scope, rate limits and data-loss rules, so an invoice agent with a valid NetSuite token still can’t change a vendor’s bank details. Calls that fail are refused before they reach the server; responses are inspected on the way back. Start in shadow mode to see what would be blocked, then enforce.

MCP gateway · decisionsenforce + shadow
  • zendesk.get_ticketSupport CopilotALLOW
  • snowflake.run_queryData Pipeline AgentALLOW
  • github.list_commitsDeploy BotDENY · RATE_LIMIT_AGENT_SERVER
  • salesforce.export_reportmarketing-gptWOULD_DENY · AGENT_INACTIVE

Illustrative. The last call ran in shadow mode: recorded as WOULD_DENY, not blocked, so you see the impact before enforcing.

One identity model

Humans and agents, one lifecycle.

Agents get the same joiner, mover and leaver workflows as people, plus an accountable owner, approvals, access reviews and separation-of-duties checks. When someone leaves, the agents they own are deactivated or reassigned in the same sweep.

Lifecycle · joiner / mover / leaverhumans + agents
  1. eventleaver · Dana Lee · Finance
  2. ownsInvoice Triage Agent, AP Reconciliation Bot
  3. agentsdeactivated with their owner, or reassigned
  4. accessNetSuite, Gmail, Snowflake → revocation queued
  5. auditeach step recorded on the audit trail

Illustrative. When a person leaves, the agents they own leave with them — one sweep, not two systems to reconcile.

Tamper-evident audit

Every decision, provable.

Allowed and denied calls, approvals and lifecycle changes land on one hash-chained, append-only audit trail. Evidence for an auditor is a query, not a quarterly project.

Audit chainappend-only
  • #1281 ACCESS_REQUEST_APPROVEDprev 7c1e…a90f → hash b42d…1c07
  • #1282 AGENT_SUSPENDEDprev b42d…1c07 → hash e9a0…5d33
  • #1283 AGENT_DEACTIVATEDprev e9a0…5d33 → hash 03fb…88e2

chain verified · no gaps, no edits

Illustrative. Each entry carries the hash of the one before it, so a deleted or altered entry breaks the chain visibly.

Plain-English answers

Ask, within your own permissions.

Ask who can do what, why they have it and when it was last used. Answers come from your live data and never exceed your own role; anything that changes access still goes through approval.

What do you want to know?

  • Who can change vendor bank details in NetSuite?
  • Which agents haven’t used their access in 30 days?
  • Grant me admin on github-prod.

you > Who can change vendor bank details in NetSuite?

3 identities hold that permission: 2 people in AP, 1 agent.

· the agent was granted it 41 days ago and has not used it since

· suggested: revoke — opens an access review for its owner

Illustrative. On your tenant, answers come from your live data and never exceed your own role.

Everything in the platform today

Enforce

7 controls on every call

  • Inline MCP gateway on every agent tool call watch it →
  • Per-tool scope checks
  • Per-agent and per-server rate limits
  • Built-in and custom data-loss rules
  • Shadow mode before enforce
  • Human approval holds for risky actions
  • Short-lived, revocable agent tokens

Govern

6 lifecycle workflows

  • Joiner, mover, leaver for humans and agents watch it →
  • Accountable owner for every agent
  • Access requests and approval workflows watch it →
  • Access reviews and certifications
  • Separation-of-duties detection watch it →
  • Break-glass access with review

Prove

4 kinds of evidence

  • Hash-chained, append-only audit trail
  • SOC 2 evidence packs on demand watch it →
  • Controls mapped to compliance frameworks
  • Personal-data export and erasure requests

Integrate

5 kinds of integration

  • 12 connectors: AWS, Entra, GCP, Okta, GitHub and more
  • SCIM provisioning
  • REST API with a versioned OpenAPI 3.1 spec
  • TypeScript SDK and GitHub Action
  • Signed outbound webhooks

Platform

3 safeguards

  • Tenant isolation enforced in the database
  • KMS envelope encryption for secrets
  • Passkeys and multi-factor sign-in

Agents in production

What goes wrong when agents act on their own, and what stops it.

AtmanID decides on each tool call that passes through its gateway, so these are caught while the agent is running, not in next quarter's review. It sits alongside Okta, Entra and SailPoint: keep them for sign-in and governance, add AtmanID where agents act. New rules can start in warn mode.

  • An agent stuck in a loop hammers an API, or issues refund after refund.

    Each agent has a call-rate limit, overall and per server, and a policy can cap what it moves through a tool each day.

    support-agent → stripe.create_refund { amount: 40 }

    daily refund cap: 4,980 of 5,000 used

    RefusedINTENT_DENY

  • Instructions hidden in a tool description or another agent’s output steer an agent off course.

    A steered agent still reaches only the tools it was granted. Poisoned tool descriptions are flagged, and so is data read from one server and sent to another.

    research-agent ← billing.get_invoice · response carries a card number

    research-agent → mail.send_email

    FlaggedTAINTED_SESSION

  • A payment, a deletion or a permission change should not happen without a person saying yes.

    Policies hold the call, on the action or its arguments, until an approver decides. One approval releases one call.

    billing-agent → stripe.create_refund { amount: 1200 }

    policy: refunds over 500 need approval

    Heldwaiting for an approver

  • On the way to an answer, an agent reaches for a tool it was never meant to use.

    Every call through the gateway is checked against the agent’s scope and intent policy, and refused before it reaches the tool.

    deploy-agent → github.delete_repo

    intent policy: deploy-agent may not delete

    RefusedINTENT_DENY

  • An agent goes beyond what it was given, or its credentials are in the wrong hands.

    Scope probing, dormant keys bursting into activity and delegation laundering raise detections. The response holds, suspends or kills the agent: as a recommendation first, then on its own.

    detection: idle for weeks, then a burst of calls

    kill switch: tokens revoked · next call

    RefusedTOKEN_REVOKED

Try it

You are the Invoice Triage Agent. Make a call.

Its owner granted it read access to NetSuite and nothing else. Pick a call and watch the gateway decide.

AtmanID MCP gateway · ENFORCE

Pick a call to send it through the gateway.

A scripted example of real gateway checks. In SHADOW mode the gateway records each of these as a would-deny and lets the call through, so you can see what enforcement would do before you turn it on.

Early access program

Free early access for five teams running AI agents in production.

What you get

  • · Early access to AtmanID, set up with your team
  • · A direct line to the founder and a real say in the roadmap
  • · Free for the duration of the program

What we ask

  • · A named owner on your side and a short check-in every two weeks
  • · Honest feedback on what works and what does not
Apply for early access

Why AtmanID

One record for every identity — person or agent.

Every agent is tied to the person accountable for it, in the same record as that person’s own access. So the gateway knows who an agent is acting for on every call, an owner’s departure takes their agents’ access with them, and one tamper-evident audit trail answers for both. Audit-readiness becomes part of normal operations, not a quarterly project.

Ātman is Sanskrit for the true self: the one identity behind every account, key and agent session.

Built for identity admins, CISOs, platform engineers and auditors.