Frequently Asked Questions

No. AgentMink is a stateless edge proxy. Prompts and model responses stream in-flight over TLS 1.3 between your application and the upstream provider. AgentMink logs only operational metadata (token counts, latency in milliseconds, model used, HTTP status, and timestamp) to your workspace ledger. The actual textual content of prompts, images, and completions is never written to disk or database.

Absolutely not. AgentMink does not train models. Furthermore, when you pass your own provider keys (BYOK), all upstream zero-training and data protection commitments from OpenAI, Anthropic, Google, and other vendors remain 100% active.

Because AgentMink is strictly stateless and does not persist payloads, PII and PHI pass in-flight over encrypted edge tunnels without being recorded. For healthcare and enterprise teams with strict compliance mandates, we offer Business Associate Agreements (BAAs) and dedicated VPC routing tunnels on Enterprise tiers.

AgentMink operates across 300+ Cloudflare edge data centers worldwide. Requests terminate at the closest edge node to your server or user, and are routed over private fiber backbones directly to the selected model provider’s nearest region for minimum hop latency.

Yes. AgentMink complies with GDPR requirements for data in transit and processor obligations. We do not store EU customer prompt data, and all telemetry records contain anonymized metadata. SOC 2 Type II audit reports are available under NDA for Enterprise plan customers.

All connections to gateway.agentmink.com enforce TLS 1.3 (with fallback to TLS 1.2) with modern cipher suites and strict HTTP Strict Transport Security (HSTS). Unencrypted HTTP traffic is rejected immediately with a 301 redirect.

Yes. You can wipe all historical telemetry ledger records or delete your workspace account permanently from your dashboard settings. Account deletion immediately purges all encrypted keys and metadata from our distributed edge storage.

All upstream requests made by AgentMink edge workers to model providers use strict mutual TLS certificate verification directly against official provider hostnames (api.openai.com, api.anthropic.com, etc.), preventing DNS hijacking and proxy interception.

Cost-First routing evaluates active per-token pricing across all connected providers every 6 hours and dynamically selects the lowest-cost model meeting your minimum reasoning quality threshold. This allows you to process batch extraction and summarization for up to 70% less than frontier list prices.

Quality-First routing evaluates benchmark performance across standardized reasoning, mathematics, and coding benchmarks (MMLU, HumanEval, MATH). It routes prompts to the highest-performing frontier model and automatically fails over to equivalent top-tier backups if rate limits occur.

Latency-First routing maintains a rolling real-time map of p95 latency and Time-to-First-Token (TTFT) across global provider data centers. Prompts are dispatched to the lowest-latency healthy endpoint available at that exact millisecond.

If an upstream provider returns an HTTP 429, 500, 502, 503, 504, or a network socket timeout before generation begins, AgentMink catches the error in under 50ms and seamlessly re-executes the prompt against your designated fallback model without throwing an exception to your client.

If a connection failure occurs before the first chunk is sent, AgentMink re-routes to a fallback provider and streams seamlessly. If a provider connection drops mid-stream after partial tokens have been delivered, the stream terminates cleanly with an EOF event to prevent duplicate token insertion.

Yes. You can pass the x-mink-strategy: cost_first | quality_first | latency_first HTTP header or pass a specific x-mink-fallback-model header on individual API requests to override workspace default rules.

Pre-dispatch gateway latency is under 5ms. Because AgentMink runs on distributed edge isolates directly adjacent to cloud provider backbones, the sub-50ms latency savings from choosing closer provider data centers often makes requests faster than calling distant APIs directly.

The edge pricing matrix synchronizes automatically every 6 hours via distributed cron workers, ensuring your cost arbitrage calculations always reflect current token rates.

BYOK allows you to store your own provider API keys (OpenAI, Anthropic, Gemini, Groq, DeepSeek) in AgentMink. Requests route directly against your existing billing agreements with those providers with zero platform markup.

Every vendor secret is encrypted using AES-256-GCM with a unique per-workspace initialization vector before storage. Keys are decrypted strictly in-memory inside isolated edge execution sandboxes during request dispatch and are never logged or exposed in client bundles.

No. Keys are encrypted with master encryption keys that prevent plain-text viewing in database tables or internal management consoles.

No. AgentMink is a 100% Bring Your Own Key (BYOK) gateway. You connect your direct provider API keys (OpenAI, Anthropic, Google Gemini, Groq, DeepSeek, AWS Bedrock, etc.) to your secure encrypted key vault. This ensures you retain 100% of your volume discounts and pay zero token markups.

Yes. You can add, update, rotate, or revoke vendor keys instantly in your dashboard settings. Key updates take effect globally across all edge nodes within 0 milliseconds.

If your vendor key returns an upstream quota error (insufficient_quota), AgentMink flags that key in cooldown, dispatches an alert, and automatically routes subsequent requests through your backup keys in your BYOK vault.

AgentMink calculates exact costs based on the prompt tokens and completion tokens reported by upstream model providers, multiplying usage by official per-million token rates to the fifth decimal place.

You can set hard daily spend limits and monthly request quotas in your dashboard. When your application reaches 80% and 100% of its budget, AgentMink dispatches webhook notifications. Once the hard cap is reached, further requests return 402 Payment Required to prevent surprise bills.

Your workspace rate limits and spending caps immediately intervene. If an unauthorized spike occurs, the gateway blocks excess traffic within 1 request of reaching your threshold, saving thousands of dollars in unexpected charges.

None. On Starter, Scale, and Enterprise plans, BYOK execution is 100% zero-markup. You pay only your flat AgentMink subscription and your direct provider rates.

We accept all major credit cards (Visa, Mastercard, American Express), Apple Pay, and Google Pay through Stripe. Enterprise plans can also pay via ACH transfer, wire, or annual purchase orders.

Yes. Plan changes take effect immediately, with prorated billing calculated automatically. You can manage subscriptions and download VAT invoices anytime through the self-service billing portal.

Migration requires changing only two lines of code: update your baseURL to https://gateway.agentmink.com/v1 and replace your API key with your AgentMink token. All OpenAI client libraries (Python, Node.js, Go, PHP, Java) work out of the box.

Yes. AgentMink passes through tools, tool_choice, and response_format parameters to supporting upstream providers with full fidelity.

Yes. Simply configure createOpenAI({ baseURL: ‘https://gateway.agentmink.com/v1’, apiKey: process.env.AGENTMINK_API_KEY }). All useChat, useCompletion, and streamText hooks are fully supported.

Yes. Pass openai_api_base=’https://gateway.agentmink.com/v1′ and your AgentMink key to any LangChain ChatOpenAI or LlamaIndex LLM instance.

Paste your incoming webhook URL into workspace settings. AgentMink dispatches rich JSON alerts whenever a spend cap is triggered, a provider 5xx failover occurs, or an error rate spike exceeds 5%.

Yes. On Enterprise and Scale plans, you can register custom OpenAI-compatible base URLs and private endpoints to route traffic across your internal infrastructure.

We guarantee a 99.99% monthly gateway uptime SLA on Scale and Enterprise plans, backed by globally distributed multi-region edge nodes and automated provider health checks.