Skip to content
AI platform

AI that is a platform capability, not a bolt-on

Every plugin can call AI through one gateway. You pick the model, the budget and the escalation rules — and switching provider never touches anything you built.

Scripted example — not a live model

Where is my order #48120?

AI agent

It shipped yesterday and is out for delivery today. Want the tracking link?

In production this runs through the AI gateway on whichever provider the tenant has configured — Claude, OpenAI or Gemini — with the knowledge base and tools you allow.

Vendor neutrality

No plugin ever imports a model SDK

That single rule is why you can change provider with a settings toggle rather than an engineering project.

Claude

Anthropic

Long context, careful reasoning, tool use

GPT

OpenAI

Broad ecosystem, structured outputs, vision

Gemini

Google

Multilingual, cost-efficient at volume

Yours

Future provider

One adapter, zero plugin changes

plugin.ts
// Any plugin can request AI. It asks for a PURPOSE, never a vendor.
const result = await ctx.ai.complete(
  {
    purpose: "classify",              // chat | summarise | classify | extract | …
    messages: [
      { role: "system", content: "Classify the customer intent." },
      { role: "user", content: message.body },
    ],
    tools: [lookupOrderTool],         // contributed by another plugin
  },
  scope,                              // tenant scope — always required
);

// The gateway picked the provider from tenant policy and emitted
// ai.response_generated with real token counts and cost.
result.provider;  // "anthropic"
result.usage;     // { inputTokens, outputTokens, costMinor }

How routing decides

  1. 1An explicit provider on the request, if one was given.
  2. 2The tenant’s policy for that purpose — classification can use a cheaper model than customer-facing chat.
  3. 3The tenant’s default provider.
  4. 4Any registered provider, so a misconfiguration degrades rather than fails.

Adding a future provider means writing one adapter that implements the provider contract. No plugin changes, no core changes, no migration.

Capabilities

What the AI actually does here

Not a chat box bolted to the side — capabilities that appear inside the modules where the work happens.

AI agentsAI

System prompt, knowledge, tools and a defined escalation path to a human.

Knowledge baseAI

Documents and site crawls, answered with source attribution rather than confident invention.

Tool callingAI

Look up an order, check stock, book a slot — using tools your plugins expose.

TranslationAI

Inline, both directions, with the original preserved for the record.

Sentiment scoringAI

Surface conversations heading toward escalation before they get there.

Document AI & OCRAI

Extract structured fields from images and PDFs customers send you.

Runtime

The unglamorous parts that decide reliability

Streaming and tool calling get demoed. Retries, budgets and cost attribution are what you live with.

Streaming

Tokens render as they arrive, so a long answer never looks like a hang.

Structured outputs

Ask for JSON and get JSON, validated before it reaches your workflow.

Retry & fallback

Transient provider failures retry; persistent ones fall back to your second choice.

Rate limiting

Per-tenant ceilings so one account cannot exhaust a shared key.

Cost tracking

Every response records input and output tokens and the exact cost.

Prompt versioning

Prompts are versioned artefacts — change one and you can see what changed.

Getting started

Four steps to a grounded agent

Most teams have something answering real questions the same afternoon.

  1. 1

    Add your keys

    Bring your own Anthropic, OpenAI or Google keys. They are stored write-only and never returned by the API.

  2. 2

    Set routing

    Choose a default provider per purpose — chat, classify, extract, translate — and a monthly budget ceiling.

  3. 3

    Ground the agent

    Upload documents or crawl your site. Answers cite their source so an agent can be checked.

  4. 4

    Give it tools

    Every installed plugin can lend the agent capabilities, always scoped to your tenant.

Safety

An agent should know what it must not do

Guardrails are configuration, not prompt-engineering folklore. Tools that change state are marked as mutating and can require confirmation. Tools always run inside the tenant scope of the conversation. And every tool invocation is recorded.

  • Mutating tools flagged, and confirmable before they run
  • Tenant scope passed to every tool — not a prompt instruction
  • Confidence threshold with automatic human handover
  • Full audit of which tool ran, with what arguments, and whether it succeeded
  • Monthly budget ceiling that refuses rather than overspends

Agent configuration

Live
ProviderClaude (Anthropic)
Purpose routingchat → Claude · classify → Gemini
Knowledge base4 documents · 312 passages
Toolslookup_order · search_products · book_appointment
Confidence threshold0.72 — below this, hand to a human
Monthly ceiling₹ 25,000

Every field here is a setting on the plugin, editable per tenant and recorded in the audit log when it changes.

Questions

What teams ask about AI

Building on it? The plugin SDK shows how to contribute your own AI tools.

Bring your own keys and see it on your data

On the call we will point an agent at your actual knowledge base and show you where it is confident, where it is not, and what it costs.

No credit card required · 14-day trial · Cancel anytime