Works with the AI Agents you already use

    Claude CodeClaude CodeCursorCursorCodex CLICodex CLIDeepSeekDeepSeekMuseMuse+22 more

    AI Gateway Architect

    1

    Design a vendor-neutral AI gateway that routes each workload through the right eligible model while controlling policy, privacy, reliability, latency, and cost.

    $9.99

    Secure checkout via Stripe

    Money-back guarantee

    Converts to your local currency at checkout

    Security scanned
    AI Gateway Architect

    AI Gateway Architect

    Example session with this skill installed

    We run a multi-tenant SaaS product and are adding a model gateway. Current details: workloads include support chat, document extraction, and code assistance; providers are not yet selected; customer data ranges from public to confidential; we need EU residency for some tenants; chat is streamed; extraction requires JSON; budgets and per-tenant usage reporting are required. We have no measured traffic or latency SLOs yet. Propose an initial vendor-neutral architecture, the intake questions and validation register, a routing policy, registry fields, reliability controls, and a staged pilot. Clearly mark assumptions and do not invent provider pricing, regional availability, or model capabilities.

    • Read your context and instructions
    • Compiled the ai gateway architect

    Initial design: governed multi-tenant AI gateway

    Executive summary

    Start with a policy-enforcing control plane in front of provider adapters. Resolve tenant identity and policy first, classify the workload and data next, filter ineligible routes, then optimize among verified eligible candidates. Keep provider credentials server-side. Treat streaming, structured extraction, and code assistance as separate workload profiles. No provider or model has been selected or verified in the supplied information.

    Reference request path

    1. Authenticate the application and resolve tenant, environment, plan, BYOK setting, and policy version.
    2. Validate request shape and classify workload, modality, data sensitivity, region, tool needs, and output constraints.
    3. Apply hard eligibility rules: permitted provider/model, residency, contract, tenant, safety, modality, schema, and retention.
    4. Check health, quota, concurrency, and capacity; estimate cost only when current pricing and token assumptions are available.
    5. Rank eligible routes against workload-specific quality, latency, reliability, cost, and evidence confidence. Return an explainable route decision.
    6. Execute with bounded timeouts and retries; preserve stream semantics. Record route, policy version, outcome, latency, and usage metadata under the approved retention policy.
    7. Evaluate quality and operational signals; update health and routing only through versioned changes.

    Workload policy draft

    WorkloadRequired controlsCandidate-selection notes
    Support chatTenant policy, sensitivity, region, streaming, safety, budgetValidate EU eligibility per tenant; do not route confidential data by default without an approved provider policy
    Document extractionData class, region, schema validation, completeness checksRequire JSON/schema capability verified for the chosen model; validate outputs before downstream use
    Code assistanceData classification, repository boundary, tool policy, retentionTreat proprietary code as confidential unless classified otherwise; isolate tenant context

    Registry fields

    Maintain canonical model and provider IDs, endpoint and credential references, modalities, context and output limits, tool/schema/streaming support, verified regions, data-handling terms, pricing/quota evidence, health, deprecation state, tenant eligibility, evidence source, and last-verified timestamp. Keep aliases versioned (for example, support_chat_eu and extraction_json_eu) rather than hard-coding provider model IDs in applications.

    Reliability and cost controls

    Use request deadlines, bounded retry budgets, idempotency-aware retry rules, circuit breakers, concurrency and quota limits, and explicit fallback compatibility checks. Do not silently downgrade region, privacy, output schema, or quality requirements. Meter usage by tenant, application, workload, route, and policy version; separate budgets from provider rate limits and capacity. Disable prompt/response caching until privacy, freshness, tenant-isolation, and provider-term rules are approved.

    Assumptions and validation register

    • Assumption: the gateway can identify tenant and workload before routing. Validate authentication and identity propagation.
    • Unknown: EU residency meaning (processing, storage, support access, or all three). Confirm legal and contractual requirements with the responsible team.
    • Unknown: provider/model eligibility, JSON guarantees, streaming behavior, pricing, rate limits, regions, and data retention. Verify current authoritative documentation and contracts before implementation.
    • Unknown: traffic, concurrency, payload sizes, latency objectives, and quality baselines. Instrument a representative pilot before setting thresholds.
    • Unknown: BYOK and customer audit needs. Determine key ownership, rotation, and audit access model.

    Pilot and next steps

    Instrument one low-risk workload with synthetic or approved data; collect latency, errors, schema validity, quality review, and usage by tenant. Define acceptance thresholds with product, security, and finance stakeholders. Run shadow evaluation before traffic shifting, then use a tenant-limited canary with rollback criteria. No systems or documents were inspected and no provider performance or savings are claimed.

    Connects securely to your tools. The creator never sees your data.

    About this skill

    AI Gateway Architect

    Design, review, troubleshoot, or modernize the control plane between your applications and AI providers. This skill turns a collection of model endpoints into a governed platform with explicit workload requirements, tenant policies, capability checks, explainable routing, reliability controls, and measurable operations.

    What it helps you design

    • Workload and data-classification intake, including modality, quality, latency, region, tenant, and fallback needs.
    • Provider and model registries with evidence timestamps, capability metadata, health, policy attributes, and versioning.
    • Eligibility-first routing: enforce privacy, capability, region, contract, tenant, and safety requirements before comparing cost or speed.
    • Resilience controls such as timeouts, circuit breakers, bounded retries, fallback rules, streaming behavior, and capacity checks.
    • Budgets, quotas, usage metering, cost attribution, caching rules, evaluations, observability, and migration plans.

    Typical deliverables

    A current-state assessment, target architecture, request-routing sequence, policy matrix, registry schema, failure-mode analysis, rollout plan, evaluation approach, and decision log. Recommendations distinguish verified facts, supplied evidence, assumptions, and open questions. Provider prices, model capabilities, limits, availability, regions, and performance are never invented; time-sensitive details must be verified against current authoritative sources.

    How to use it

    Describe the applications, workloads, providers, tenancy model, data classifications, required regions, quality and latency targets, and operating constraints. If information is missing, the skill proceeds with clearly labeled assumptions and a validation register. It does not claim to inspect source code, traces, contracts, dashboards, or provider documentation unless you supply them or connect an available tool.

    Guardrails

    Sensitive workloads must not be routed to an ineligible provider. Provider credentials remain server-side and tenant-isolated. Prompt and response retention is opt-in and policy-bound. Caching must respect freshness, privacy, tenant boundaries, and provider terms. Retries must not multiply non-idempotent side effects. A cheaper or faster model is not automatically an acceptable substitute.

    Find more skills like this on the Claude Skills Marketplace, the Codex Skills Marketplace and the Gemini Skills Marketplace, or browse Claude Skills for LLM Ops.

    This skill produces architecture and planning guidance; it does not build, deploy, secure, or operate a gateway. It cannot verify provider capabilities, prices, quotas, regions, contracts, or data-handling terms unless current authoritative evidence is supplied or accessible. Recommendations depend on accurate workload, tenant, privacy, and traffic inputs. Latency, quality, cost, availability, and savings are not guaranteed. Legal, regulatory, and security decisions require review by qualified owners. No credentials should be pasted into prompts. Generated configurations and policies require testing before production use.

    How to use

    1. 1

      Connect your agent to the Agensi MCP

      Copy the prompt and paste it into Claude, ChatGPT, Cursor or any agent that supports MCP. It sets up the connection for you. Manual setup

    2. 2

      Sign in to Agensi

      Your agent shows a sign-in link. Sign in and click Allow.

    3. 3

      Ask for what you need

      Your agent loads this skill and uses it on your task. It always asks before buying a skill.

    Reviews

    No reviews yet

    Be one of the first to try it. Every listed skill goes through the trust checks below.

    Security scanned

    Checked by our 8-point scan before listing

    Fresh listing

    Recently published to Agensi

    Money-back guarantee

    Doesn't do what its page promises? You get your money back.

    Trust & safety

    Security scan

    Security scan passed 1 month ago

    Permissions

    Read Files
    Write Files
    Browser
    Network Access

    Read only the files and configuration paths the user explicitly supplies or selects; declare approved paths before inspecting a repository. Write only requested design documents or diagrams to a user-approved output location, and confirm before overwriting. Browser and network access are optional and used only when available and authorized to verify current provider documentation or terms. Never collect, reveal, or transmit provider credentials, prompts, or tenant data. No provider call, deployment, policy change, or routing action is performed. Distinguish inspected evidence from assumptions.

    Listed1 month ago

    What's inside

    Frequently Asked Questions

    The AI Gateway Architect helps you move away from hardcoded model endpoints by designing a centralized control plane that automatically routes requests based on your specific requirements for privacy, latency, cost, and model reliability.

    This skill is vendor-neutral and designed to work with any model provider or orchestrator, including OpenAI, Anthropic, Azure AI, AWS Bedrock, and self-hosted models via vLLM or TGI.

    You will receive a comprehensive architectural blueprint, including request-routing logic, a provider registry schema, a failure-mode analysis, and a detailed rollout plan tailored to your specific tenancy and data requirements.

    No, the skill uses an 'eligibility-first' framework where privacy, regional compliance, and capability requirements are verified before any routing decisions based on cost or speed are made.

    The skill identifies specific performance targets and resilience controls, such as circuit breakers and bounded retries, to ensure your AI services remain stable even if a specific provider experiences latency spikes or outages.

    New in LLM Ops