The AI gateway
Know which way the AI wind blows.
Vane sits between your company and every AI provider. It sees each request, applies your rules, cleans out sensitive data, controls cost, and proves it all to auditors.
Runs in your network · Content capture off by default · Monitor mode first
Figures from the Vane demo workspace: a 40-person example company.
The problem
Your firewall sees an address. AI risk lives inside the request.
Every person with a key, every agent and every copilot sends company data to an outside model. Network tools can't read it. Vendor consoles only see their own slice.
What is an AI gateway? A single point that every AI request passes through on its way to a model provider. Because it sees each call, it can apply company rules, remove sensitive data, control cost and record what happened. Vane is an AI gateway.
| In one AI call | Firewall / proxy | Vane |
|---|---|---|
| Who sent it | An IP address | A person, team or agent, with an owner |
| What is inside | Encrypted bytes | Customer IDs, card numbers, keys, found and removed |
| Which model and tools | A hostname | Model, MCP server and tool, checked against an allowlist |
| What it costs | Nothing | Tokens, price and budget, per call |
| What happened | A connection log | A decision, with a reason, in a tamper-evident trail |
The platform
See it. Steer it. Prove it.
Six controls in one gateway. Pick one to watch it work on a real request.
Policies in plain English
Write the rule the way you would say it. Vane turns it into a policy, shows you what it would have caught last week, and starts in Monitor mode until you switch it on.
- Scope by team, agent, person, model or MCP server
- Block, redact, warn or ask for approval
- Test against sample text before it goes live
Sensitive data never leaves
Vane finds sensitive values in the request and replaces them before the call goes out. The model still gets enough to do the work. The values are never stored.
- Cards checked with Luhn, IBANs with mod-97, Israeli IDs with the check digit
- API keys, private keys, JWTs and passwords
- Responses scanned on the way back too
Budgets that hold
Every call has a price and an owner. Vane starts with a cheaper model and moves up only when the request needs it. It caches long prompts and reuses answers to repeated questions.
- Budgets for the organization, a team, an agent or a person
- Each saving set to Monitor, On, Off or Inherit
- Lock a setting, or hand it to the budget owner
Every agent, on one map
See which agents run on your network, who made them, which models and MCP tools they call, and which agents talk to each other. An agent without an owner is flagged until someone claims it.
- Owner, team and network segment for each agent
- MCP server and tool allowlists
- Block an agent in one click
Ask first, from your phone
Some calls need a person to say yes. Vane holds the request, sends the approver a push notification, and lets the call through the moment they approve.
- Approve once, for a day, or for good
- iPhone app with push notifications
- Every decision lands in the audit trail
Wants to send 412 customer rows to claude-sonnet through the MCP tool crm.export.
Evidence your auditor accepts
Each policy decision and admin action is written to a hash-chained trail. Change one entry and the chain breaks. Reports map the evidence to the controls you are audited on.
- Mapped to SOC 2 and ISO/IEC 27001 Annex A
- Daily report as PDF, short video or podcast
- Request IDs returned to every app
ISO 5.34Privacy and protection of PIIEvidenceISO 8.2Privileged access rightsEvidenceSOC 2 CC7.2Monitoring of system activityEvidenceHow it works
Live in under an hour, without touching your apps' code
- Step 1
Run Vane in your network
One Docker deployment, inside your network. Your data plane stays yours.
# one command docker compose up -d - Step 2
Point your AI traffic at it
Change the base URL in your apps, agents and SDKs. Keys, prompts and responses work as before.
base_url=https://vane.acme.internal - Step 3
Watch, then steer
Policies start in Monitor mode. See what they would do for a week, then switch them on.
policy.mode = "monitor"
By role
One gateway, four owners
Eight built-in roles, from viewer to admin. Each team sees what it owns.
CISO
Control what leaves
Sensitive data removed at the edge. Unknown models and tools blocked. Shadow AI found by the network sensor.
CFO & FinOps
One AI number
Spend by team, agent and person, with budgets that hold and savings you can verify.
CTO & Platform
Ship AI faster
One endpoint for every model. Streaming pass-through, fail-open by default, no code changes.
Compliance
Prove it
A tamper-evident trail and evidence packs mapped to SOC 2 and ISO/IEC 27001.
Security
“Do you keep our prompts?” No.
Vane records the decision, the model, the cost and the owner. Content capture is off by default. Detected values are never stored.
In your network
The gateway runs where your data lives. Nothing loads from outside sites.
Fail-open by default
If the gateway stops, your AI apps keep working. Choose fail-closed per policy.
Strong sign-in
Expiring tokens, constant-time checks and lockout after repeated failed sign-ins.
Scoped roles
Admin, approver and viewer kept apart. Break-glass access is logged by name.
Compliance and the iPhone app
The answer is in your pocket before the auditor asks
Vane sends each owner a daily report: as a PDF, a two-minute video or a short podcast. Approvals and budget alerts arrive as push notifications on the Vane iPhone app.
- Evidence mapped to SOC 2 and ISO/IEC 27001 Annex A controls
- Privacy controls aligned to ISO/IEC 27701
- Your auditor certifies your company; Vane supplies the evidence
FAQ
Questions security and platform teams ask us
What is an AI gateway?
An AI gateway is a single point that every AI request passes through on its way to a model provider. Because it sees each call, it can apply company rules, remove sensitive data, control cost and record what happened. Vane is an AI gateway.
How is Vane different from an API proxy or a firewall?
A proxy forwards traffic and a firewall sees addresses. Vane reads the AI request itself: the model, the person or agent behind it, the tools it calls and the sensitive data inside it. It decides per request whether to allow, redact, block or ask for approval.
Does Vane store our prompts?
No, not by default. Content capture is off. Vane records the decision, the model, the cost and the owner. Detected values such as ID or card numbers are never stored. If you turn capture on, content is encrypted.
Will Vane slow down our AI apps?
Vane streams responses straight through and runs its checks inside a strict latency budget. The default is fail-open, so an outage of the gateway never stops your AI apps; you can choose fail-closed per policy.
How long does setup take?
Usually under an hour. Run Vane in your network with Docker, then change the base URL in your apps and agents to point at Vane. Policies start in Monitor mode, so nothing is blocked until you decide.
Which sensitive data does Vane detect?
Email addresses, phone numbers, card numbers (Luhn-checked), IBANs (mod-97 checked), US Social Security numbers, Israeli ID numbers (check digit), API keys, private keys, JWTs and passwords. Each detector is tuned for precision so it does not break real work.
How does Vane cut AI cost?
Three switches: prompt caching, answer reuse for repeated questions from the same person, and lower-model routing that starts with a cheaper model and moves up only for complex requests. Each can be set to Monitor, On, Off or Inherit for the organization, a team, an agent or a person.
Does Vane help with SOC 2 and ISO 27001?
Yes. Vane keeps a hash-chained audit trail of every admin action and policy decision, and maps its evidence to SOC 2 and ISO/IEC 27001 Annex A controls such as 5.34 (privacy and PII) and 8.2 (privileged access). Your auditor still certifies your company; Vane supplies the evidence.
Contact
See Vane on your own traffic
A 30-minute demo on the example company, then a week in Monitor mode on yours. We reply within one business day.
info@vanesecurity.com