LLM Gateway
One endpoint for every model you want to use. Point any OpenAI SDK, LangChain, or editor extension at ApiSpi and get automatic connector tools, multi-provider routing, and governance guardrails — with no code changes on your side.
What's Included
- ✓OpenAI-compatible chat completions endpoint
- ✓Streaming and non-streaming responses
- ✓Your connected tools, auto-injected as functions
- ✓Bring-your-own-provider routing
- ✓Governance guardrails on every request
- ✓Real-time SIEM/webhook export of governance events
- ✓MCP server + A2A agent endpoints included
- ✓Usage tracked automatically, no separate metering
Quick Start
Built For
Developers wiring AI into internal tools and teams standardising on one endpoint across providers — plus Chief AI Officers who need governance over all of it. See who it's for below.
ApiSpi in 2026
ApiSpi is an AI agent marketplace and governance platform: a catalogue of ready-made agents, the connectors that give them real work to do, and the policy engine that keeps every request accountable. The LLM Gateway is the layer underneath all of it — the same agents, connectors, and governance your team already uses in Aria, exposed as one programmable API.
Ready-made agents
Spanning government & procurement, security & compliance, finance, healthcare, real estate, customer support, and more — subscribe and go, or build your own on the same platform.
Connectors
CRM, communications, dev tooling, finance and ERP systems, government data, and eleven AI provider engines — each one available to your agents and, through the gateway, to your own code.
Governance policy
Per-organisation policies, Microsoft Entra ID single sign-on, an AI firewall, and an emergency stop scoped to your organisation — introduced through 2026, enforced everywhere your team touches AI, gateway included.
Ways in, one engine
OpenAI-compatible chat completions, an MCP server for tool-calling clients, and an A2A endpoint for agent-to-agent delegation — all backed by the same connector tools and policy engine.
How It Works
Get an API key
Generate a gateway key from your ApiSpi profile in seconds — no separate signup.
Point your client at ApiSpi
Set your OpenAI SDK's base URL to apispi.com/api/gateway/v1 — everything else stays the same.
Requests run through your connectors
Your active connectors are auto-injected as tools and executed server-side — no orchestration code needed.
Governance applies automatically
Budgets, keyword guardrails, and injection detection run on every call — nothing extra to configure.
Bring Your Own Provider
Inference routes through whichever provider connector you have active — Anthropic, Google Gemini, SCX AI, Argyll Data, Mistral AI, DeepSeek, Groq, Grok (xAI), OpenAI, Z.ai, or Qwen (Alibaba Cloud) — so you're never locked into one vendor. No active connector? Requests fall back to the shared ApiSpi platform key automatically.
Anthropic
Claude Opus 4.8, Opus 4.7, Sonnet 4.6, Haiku 4.5
Google Gemini
Gemini 2.0 Flash, Flash Lite, 1.5 Pro
SCX AI
Default OpenAI-compatible engine
Argyll Data
OpenAI-compatible engine
Mistral AI
Mistral Large, Small, and more
DeepSeek
DeepSeek-V3 chat, DeepSeek-R1 reasoner
Groq
LPU-accelerated Llama, Kimi, and more
Grok (xAI)
Grok 4 and earlier generations
OpenAI
GPT-4o, GPT-4o-mini, o1, o1-mini
Z.ai (GLM)
Zhipu's GLM model family
Qwen (Alibaba Cloud)
Qwen-Max, Qwen-Plus, Qwen-Turbo
Governance, Built In
Every gateway request — including tool-calling passthrough from clients like Cline or Roo — runs through the same governance your team already configures for Aria.
Token Budgets & Access Windows
Daily, weekly, and monthly limits, plus time-of-day access windows, are enforced on every call.
Keyword & Injection Guardrails
Blocked-keyword scanning and prompt-injection detection run on input and output, even under streaming.
Automatic Usage Tracking
Every request is recorded against your usage automatically — no separate metering or reconciliation.
MCP (Model Context Protocol)
The gateway is also an MCP server — point Claude Desktop, Claude Code, or any other MCP client at ApiSpi and it can discover and call every tool from your active connectors directly, using the same gateway API key as the OpenAI-compatible endpoint above. No separate integration to build per connector.
Connect a client
Add ApiSpi as a remote MCP server in your client's config, using a gateway key from API Keys:
One endpoint, every connector
The MCP server exposes the exact same tool engine as Aria and the OpenAI-compatible endpoint — tools/list and tools/call map straight onto your connectors, so anything you can turn on in My Connectors is automatically available to any MCP client, with the same governance guardrails applied. This costs nothing extra to maintain — any WhatsApp, Salesforce, or Stripe connector we ship as a REST integration works over MCP automatically.
A2A (Agent2Agent Protocol)
ApiSpi is also an A2A agent — other AI agents can discover Aria and delegate whole tasks to it over the open Agent2Agent protocol. Where MCP hands a client your individual tools, A2A hands it an outcome: send a task, get back a completed result produced with your connected services, under your governance policy. Agents discover ApiSpi via its Agent Card at /.well-known/agent-card.json.
Delegate a task
Any A2A client can send work with a gateway key — synchronously, or queued with blocking: false and polled via tasks/get:
Tasks, not just tools
A2A tasks run through the same engine as Aria, background tasks, and MCP: your connectors do the work, and your governance travels with every delegation — tool rules, spend limits, approval requirements, and the emergency stop all apply to external agents exactly as they do to your own team. Long-running tasks queue and report their lifecycle (submitted → working → completed) so calling agents can fire-and-poll instead of holding a connection open.
Who Is It For?
Chief AI Officers who want to govern AI activity across their organisation — not just the requests that happen to pass through one tool. Layered with your existing API gateways and WAFs, ApiSpi gives you granular control across the entire AI ecosystem: every model, every connector, every request.
Directors who need to demonstrate due diligence for protecting company systems from AI threats. Centralised guardrails, access controls, and usage records give you an evidence trail for how AI activity is governed, not just a policy document.
Granular Governance
Sits alongside your API gateways and WAFs to extend control to the AI layer specifically — budgets, guardrails, and access policy enforced on every model call, not just at the network edge.
Cost Optimisation
Automatic usage tracking and token budgets across every provider give you a single, accurate view of AI spend — and the levers to control it — instead of reconciling bills across vendors.
Future-Proof by Design
Bring-your-own-provider routing means switching or adding a model vendor is a config change, not a re-architecture — your governance policy travels with every request regardless of which model answers it.
Deployment Options
Run the gateway however fits your environment — start on our managed SaaS today, with a self-hosted appliance on the roadmap for organisations that need the workload on their own infrastructure.
SaaS
Fully managed by ApiSpi — sign up and start making gateway requests in minutes. No infrastructure to provision or patch; governance, routing, and usage tracking run on our platform.
- ✓ Zero setup — get an API key and go
- ✓ Managed upgrades and patching
- ✓ Same governance and connector tooling as every other ApiSpi surface
On-Prem Appliance
A self-hosted deployment for organisations that need the gateway running inside their own network boundary — for data-residency, air-gapped, or regulatory requirements SaaS can't meet.
- ✓ Runs entirely within your infrastructure
- ✓ Same governance policy engine as SaaS
- ✓ On the roadmap — register your interest below
Frequently Asked Questions
Do I need to change my existing code?
No. The gateway speaks the standard OpenAI chat-completions format, so any OpenAI SDK, LangChain integration, or editor extension that lets you set a custom base URL and API key works without modification.
Which model do I get?
Whichever you request. Naming a claude-* or gemini-* model routes directly to that engine; otherwise requests use your active OpenAI-compatible connector (SCX, Argyll, Mistral, DeepSeek, Groq, Grok, OpenAI, Z.ai, or Qwen) in that order, falling back to the shared platform key if none is connected.
Can I use my own tools with function calling?
Yes. If your request supplies its own tools array — as clients like Cline or Roo do — the gateway forwards it and relays tool calls straight back, still applying the same governance guardrails to input and output.
Is streaming supported?
Yes, set stream: true as usual. If output guardrails are enabled, the reply is scanned before delivery, which adds a short delay; otherwise tokens stream live.
Is ApiSpi an MCP server or an MCP client?
A server. ApiSpi hosts the MCP endpoint at /api/gateway/mcp — your MCP client (Claude Desktop, Claude Code, etc.) connects to it, the same way it would to any other remote MCP server.
What's the difference between the MCP and A2A endpoints?
MCP exposes your individual connector tools to a client that runs its own reasoning (Claude calls your Gmail tool, then decides what to do next). A2A delegates a whole task to Aria — the calling agent sends "summarise this week's customer emails" and gets back the finished result. Same connectors, same governance, same API key; the difference is who does the thinking.
Can I stream governance events into our own SIEM?
Yes. Set a webhook URL from the admin governance page and every enforcement event on gateway traffic — guardrail blocks, spend alerts, tool approvals, policy-as-code matches — is HMAC-SHA256 signed and POSTed there in real time, alongside the on-demand CSV export and posture report.