Secure
Stop sensitive data, secrets and attacks before they reach a model, and scan what comes back.
- PII: monitor, redact, tokenise or block
- Secrets and credential detection
- Prompt-injection detection
- Response DLP scanning
AI Guardrails & Control
AI, within bounds.
Protect sensitive data. Enforce policies. Keep every AI interaction within your organisation's boundaries.
- Australian and New Zealand identifiers built in
- OpenAI and Anthropic compatible
- Provider keys never leave the gateway
gateway.tokard.com.au/ui · Usage & spend
Illustrative mock-up with placeholder values. Not real customer data.
4 modesMonitor, redact, tokenise, block
7 APIsChat, responses, embeddings, messages, images, speech, transcription
8 presetsEvery feature an on/off switch
1 keyPer app. Provider keys stay inside
One endpoint, every major provider
- OpenAI
- Anthropic
- Azure AI Foundry
- AWS Bedrock
- Google Gemini
- Mistral
- Self-hosted models
Works with
Connects to the tools your organisation already uses
Model Context Protocol (MCP) servers, governed by the same keys, budgets and guardrails.
Logos are trademarks of their respective owners and are shown to indicate compatibility only; no endorsement implied. See all integrations and their status.
The problem
AI is already in your organisation. Is anyone watching?
Staff paste client files into chat tools. Developers embed provider keys in code. Bills arrive with no owner. Each model and vendor has its own rules, or none.
- Personal and client data leaves your control
- Provider keys scattered across apps and laptops
- No budget, no attribution, no audit trail
The solution
One gateway. One policy. Every model.
Route all AI traffic through Tokard. Apps use a gateway key against one familiar API. Tokard enforces policy, optimises the request and records what happened.
- Sensitive data detected before it leaves
- Provider keys held in one place
- Every request attributed, budgeted and logged
Three pillars
Secure. Optimise. Control.
Everything the gateway does falls into one of three jobs.
Optimise
Spend less per answer and stay up when providers do not.
- Safe prompt optimisation
- Exact and opt-in semantic caching
- Smart routing with explanations
- Retries and failover
Control
Know who uses which model, for what, and at what cost.
- Virtual keys, teams and budgets
- Per-key model permissions
- Audit logs and OpenTelemetry
- Organisations and customer login
Features
Everything between your apps and the model
Each capability is a switch. Turn on what you need, monitor before you enforce.
One API for every model
OpenAI-compatible and Anthropic-compatible, across providers.
Virtual keys, teams and budgets
Per-app keys, model permissions, budgets and rate limits.
PII protection, tuned for Australia
Detect Australian and New Zealand identifiers, dates of birth and card numbers.
Secrets and credential detection
Catch API keys, tokens and passwords before they leave.
Prompt-injection detection
Flag and stop instruction-override attempts.
Response DLP scanning
Scan what comes back, not just what goes out.
Safe prompt optimisation
Trim wasted tokens without changing meaning.
Response caching
Exact caching by default. Semantic caching is opt-in.
Smart model routing
Right model for the task, with a reason you can read.
Retries and failover
Keep working when a provider does not.
Spend tracking and alerts
See cost by app, team, user, model and provider.
Audit logs and OpenTelemetry
A record you can stand behind, and export.
Policy presets, every feature a switch
Start from a preset. Every control is on or off.
Sign-in: Entra for staff, invitations for customers
Staff use Microsoft Entra. Customers sign in with an invited account.
More than chat
Images, text-to-speech and transcription through the same gateway.
Organisations and customer login
Each customer organisation gets its own admins, teams, keys and budgets.
USD and AUD cost display
Switch every cost view between USD and AUD.
How it works
A request, from app to answer
Every request passes through the same checks, in the same order, whichever model it is bound for.
- Your app sends a request
Any OpenAI or Anthropic SDK, with a gateway key instead of a provider key.
- Authenticate and apply policy
Key, team, model permissions, budget and rate limit are checked first.
- Scan the request
PII, secrets and prompt-injection checks run. Content is redacted, tokenised, blocked or just monitored.
- Optimise and route
Safe prompt clean-up, cache lookup, then the best model for the job, with failover ready.
- Provider responds
The gateway holds the provider keys. Your apps never do.
- Scan the response and restore
Response DLP runs, tokens are restored, and spend and audit records are written.
Use cases
Built for sectors that cannot be casual with data
Healthcare
Keep patient identifiers such as Medicare numbers out of prompts while clinicians and admin staff get the benefit of AI.
Healthcare presetFinance
Tokenise account and tax identifiers, cap spend by desk, and keep an audit trail for review.
Finance presetGovernment
Enforce approved models only, with strict blocking policies for sensitive information.
Government presetLegal
Protect client confidences, restrict models per matter team and record every use.
High Security presetEducation
Give staff and students safe access with budgets per faculty and guard rails on personal data.
Default presetAccounting
Handle TFNs and ABNs safely, with usage attributed to clients and teams for recharging.
Finance presetPresets are starting points. They support, but do not by themselves ensure, your compliance obligations.
Integration
Change the base URL. Keep your code.
Use the official OpenAI SDK as you do now. Point it at the gateway and use a gateway key. Anthropic-style clients can use the same gateway.
- Chat, responses, embeddings, messages, images and speech endpoints
- Streaming supported
- Works with Python, Node, curl and any HTTP client
from openai import OpenAI
client = OpenAI(
base_url="https://gateway.tokard.com.au/v1",
api_key="sk-your-gateway-key", # a gateway key, not a provider key
)
resp = client.chat.completions.create(
model="gpt-4o",
messages=[{"role": "user", "content": "Summarise this client file"}],
)
print(resp.choices[0].message.content)FAQ
Questions we hear often
Do I have to change my application code?
Almost never. Tokard speaks the OpenAI and Anthropic APIs. In most cases you change the base URL and swap your provider key for a gateway key.
Which models and providers are supported?
OpenAI, Anthropic, Azure AI Foundry, AWS Bedrock, Google Gemini, Mistral and self-hosted models, all behind one endpoint. Your administrator chooses which are enabled. Beyond chat, the gateway also handles embeddings, image generation, text-to-speech and transcription.
What happens to sensitive data in a prompt?
You choose per policy: monitor it, redact it, replace it with a reversible token that is restored in the response, or block the request. Australian identifiers such as Medicare, TFN, ABN, driver licence and passport numbers are covered, as are New Zealand IRD, NHI and bank account numbers, dates of birth and card numbers.
Do my applications still need provider API keys?
No. Provider keys live inside the gateway. Applications get gateway keys that you can limit, budget and revoke.
Will optimisation change what my prompts mean?
No. Optimisation is limited to safe changes such as whitespace, JSON compaction and duplicated context, and you can see the before and after token counts. It does not remove punctuation blindly.
Does semantic caching come on by default?
No. Exact caching is available by default and semantic caching is opt-in, because similar is not the same as identical.
Where is my data hosted and where does AI processing happen?
The Tokard gateway and its data are hosted in Australia, and prompts and responses are not stored by default. AI processing happens at whichever provider you choose. OpenAI is the default provider and is offshore, unless you connect an Australian-hosted provider or model. You choose where inference runs.
How do customers sign in?
Customers sign in at the customer login with the account their organisation administrator invites, using email and password. Microsoft Entra single sign-on is used for the platform's own staff console.
Does it replace a security review?
No. It is a control layer that helps you enforce policy and gain visibility. Detection is probabilistic, so test it against your own data and keep your other safeguards.
How much does it cost?
Plans are Starter, Business and Enterprise. Pricing depends on your usage and requirements, so please contact us for a quote. Prices are quoted in AUD ex-GST.
Keep every AI interaction within bounds
Request access and we will help you set up your first policy, key and budget.