Product

Everything between your AI tools and your models.

A gateway that enforces your rules on every request, a console where you set and prove them, and governed AI for the people who don't write code.

The gateway

One endpoint. Every request handled the same way.

agentKeyRoute speaks the APIs your tools already use: OpenAI chat completions, responses and embeddings, Anthropic messages, model listing, and image, speech, rerank and moderation endpoints, with streaming. A governed key identifies the caller, its team, its workload and its governance profile, so the right rules apply automatically.

Endpoints whose content can't be inspected before it leaves, such as audio transcription, batch and file uploads, are refused, not waved through.

  • POST /v1/chat/completionsOpenAI-compatible, streaming
  • POST /v1/messagesAnthropic-compatible, streaming
  • POST /v1/responsesOpenAI Responses API
  • POST /v1/embeddingsEmbeddings for RAG
  • GET /v1/modelsWhat this key may use
  • POST /v1/images · /audio/speechMedia, same checks
  • POST /v1/rerank · /moderationsSame checks and audit

Detection & redaction

Find sensitive data, and remove it before it leaves.

Every request is scanned for health, personal, payment, secret, CUI and confidential data. Detection combines patterns, checksums (so a random 16-digit number isn't mistaken for a card) and context, such as clinical language that turns a name into health data.

  • Health (PHI): patient names, medical record numbers, dates of birth, NPI numbers, health plan IDs
  • Personal (PII): social security and passport numbers, email addresses, phone numbers, street addresses, IP addresses
  • Payment (PCI): card numbers (checksum-verified), bank account numbers and IBANs
  • Secrets: API keys, AWS access keys, private keys, tokens and passwords, always blocked by every built-in profile
  • CUI and confidential markings for government and internal documents

A workload can also declare that it always handles a data type, for example a clinical tool, so protection applies even when detection finds nothing. A prompt-injection guard and output scanning run on every request.

What the tool sent

Summarise the discharge note for patient John Carter, MRN 00482913, DOB 03/14/1961. Bill card 4111 1111 1111 1111.

What the model receives

Summarise the discharge note for patient [REDACTED:PERSON_NAME], MRN [REDACTED:MRN], DOB [REDACTED:DATE_OF_BIRTH]. Bill card [REDACTED:CREDIT_CARD].

Real output of agentKeyRoute's detection engine. Typed placeholders keep the request meaningful to the model.

Policy & routing

Rules that decide where each request may go.

Policies are ordered rules with default deny. A rule can name specific models, or describe them, such as "on-prem, then private cloud", "processing in the EU", "provider holds zero data retention or no-training terms" or "good at classification, cheapest first". Described targets resolve live, so rules keep working as providers and models change.

Compliance frameworks compile into routing constraints. Combine several, and the strictest rule wins. Every change is versioned, with snapshots and one-click rollback.

Default rules for a new organisation

  1. health-data-on-prem-first
  2. government-data-us-only
  3. personal-data-off-public-internet
  4. confidential-private-cloud
  5. everything-else-requested-model

Plus a library of 22 ready-made rules for data protection, data residency, vendor contracts, cost and quality.

Task-aware routing & fallback

Each request goes to the best allowed model for the task, with fallbacks, retries and a circuit breaker when a provider struggles.

Zero-retention routing

Through OpenRouter, sensitive requests go only to endpoints on the public zero-data-retention list, and the gateway records where each call actually ran.

Effective policy per key

For any key, see which models each type of data would reach, whether it would be redacted, or exactly why it would be blocked.

Subscription accounts

Your Claude, ChatGPT, Copilot and Gemini subscriptions, governed.

Connect AI subscription accounts directly, next to your API providers. People sign in on the vendor's own page, by pasting a code, pasting an address or using a device code. agentKeyRoute never sees the password, and the account's tokens stay inside agentKeyRoute: tools only ever get their governed agentKeyRoute key.

Connect several accounts per provider and requests rotate across them. Subscriptions are treated as public cloud with no attestations, so frameworks keep sensitive data away from them, and every request is redacted, checked, metered and audited like any other. Each request records which connected account served it.

Turned on per organisation on request. Vendors license consumer subscriptions for personal use and may suspend accounts used for automated or shared traffic.

ClaudePro · Max
ChatGPTChatGPT · Codex
GitHub CopilotCopilot subscription
Google GeminiGoogle account
  • Sign-in on the vendor's page
  • Several accounts, rotated automatically
  • Same policy, redaction and audit
  • The serving account recorded per request

Routing engine

Keeps AI traffic flowing when providers don't.

agentKeyRoute's routing engine sits behind the gateway. It handles the unglamorous work of talking to dozens of providers and accounts, and it can only use the connections each organisation's policy allows: every organisation's routing keys are restricted to its own connections.

Retries & fallback

Failed calls are retried, then passed to the next model in the chain your rules allow.

Circuit breaker

A provider that keeps failing is taken out of rotation, then tried again once it recovers.

Request queue

Bursts of traffic wait in a queue instead of failing when a provider's limits are reached.

Cooldowns

Rate-limited models and accounts are rested for a while, then return automatically.

Account rotation

Requests spread round-robin across every connected account of a provider, API keys and subscriptions alike.

Concurrency & rate limits

Set the maximum concurrent requests per provider, with built-in protection against hitting rate limits.

Prompt compression

Long prompts can be compressed before sending, to cut tokens, cost and latency.

Provider prompt caching

Uses providers' own prompt caching where they offer it, so repeated context costs less.

Live health

Each provider's success rate, latency and queue, right on its page in the console.

Guardrails & risk

A second opinion from a safety model you host.

Mark any model as a guardrail, such as Llama Guard 3, ShieldGemma or Granite Guardian, and use it to screen requests and, optionally, responses against the 14 hazard categories of the MLCommons taxonomy. You choose which categories block, and whether to fail closed or open if the guardrail is down.

Built-in profiles only ever choose your own on-premises or private-cloud guardrail, so screening can never send data somewhere the request itself couldn't go.

  • Risk scoring with record, hold or block per tier
  • Incidents opened automatically for high-severity blocks
  • Denied topics and maximum prompt size per profile
  • Blocking of images, audio, video and files for sensitive workloads
  • Kill switches to stop a workload immediately
  • A risk register with key risk indicators

Audit & evidence

Proof that the rules were followed, request by request.

Model verification

After each call, the gateway checks which model actually answered. If it isn't one the policy allowed, the request fails closed.

Hash-chained audit log

Every request is metered, recorded and chained to the one before it. An integrity check shows whether any entry was altered.

Evidence packs

19 frameworks with automated control checks and attestation. Export an evidence pack for your auditor in one click.

Cost control

See and steer what AI costs.

Every request is metered by team, workload, person and model. Set budgets per workload, enforce plan limits, and produce chargeback reports. Where a provider reports the exact cost, it replaces the price-list estimate.

Usage outside the gateway counts too: import daily Claude usage and cost per person and product from Anthropic's analytics API (aggregates only, no conversation content).

  • Encrypted response cache per organisation, profile and model chain
  • Health, card and secret data are never cached by default
  • Cache savings shown in Cost & usage
  • Workload budgets and plan limits
  • Live traffic and observability

For everyone

Approved AI instead of personal accounts.

agentKeyRoute Chat

A chat window in the console for everyone in the organisation. Messages follow the chat profile your administrators pick and are attributed to each person.

AI workspace

A coding agent with chat and code modes, in a private sandbox per person: non-root, all capabilities dropped, memory and CPU limits, an isolated network that can't reach your databases. Its settings force it to use only your governed models, with a key issued to that person. It can also create images, speech and video through the gateway.

Browser extension

For Chrome and Edge. It checks what people type into ChatGPT, Claude, Gemini, Copilot and Perplexity before it's sent. Your profile decides: allow, redact the details in the box, or block. Every check appears in your request log.

Providers

38 provider types behind one key.

Add a provider with its own key, and its models sync automatically when it's added and every six hours, typed by what they can do. Models a provider stops listing are disabled, not deleted.

Model vendors

OpenAI · Anthropic · Google Gemini · xAI Grok · Mistral AI · Cohere · DeepSeek · Moonshot AI (Kimi) · Alibaba Qwen (Model Studio) · Z.ai (GLM) · Perplexity

Aggregators & inference clouds

OpenRouter · Together AI · Fireworks AI · Groq · Cerebras · DeepInfra · SambaNova · Hyperbolic · Novita AI · Hugging Face Inference · NVIDIA NIM (hosted) · OpenCode (pay-as-you-go) · ExcelScale

Your cloud tenancy

Azure OpenAI / AI Foundry · AWS Bedrock · Google Vertex AI

Self-hosted

Ollama · vLLM · NVIDIA NIM (self-hosted) · Hugging Face TGI · LM Studio · LiteLLM proxy · Any OpenAI-compatible endpoint

AI subscriptionsdirect

Claude Pro / Max · ChatGPT / Codex · GitHub Copilot · Google account (Gemini)

See every supported provider and model →

The console

Set the rules, watch the traffic, prove compliance.

Guided setup

A live checklist in five phases, from connecting models to proving compliance, that ticks itself off as you configure.

Connect an app in three steps

Describe the tool and its data, accept the recommended profile, and get a key plus a ready-made setup snippet.

Request explorer

Every request with its full decision trace: what was found, which rule applied, where it went and what it cost.

Policy playground

Test a request against your rules before you roll a change out.

Workloads

Lifecycle, owners, pre-launch checks, EU AI Act tier and budgets for each AI use case.

Organisation controls

Users, roles, teams, two-step sign-in, two-administrator approval and SIEM or webhook destinations.

See it with your own tools.

Connect a provider or subscription and point one tool at agentKeyRoute. Once your organisation is approved, the first governed request takes a few minutes.