# 12 · Artificial intelligence

https://exo.now/docs/stack/ai

> Apps send prompts to Exo Router, Exo's smart model routing. It scores how hard each task is and picks the least costly capable model through the Convex AI Gateway.



<StackFacts dimension="12 · Artificial intelligence" platforms="[&#x22;convex-ai-gateway&#x22;, &#x22;artificial-analysis&#x22;, &#x22;arch-router&#x22;]" status="Exo Router endpoint live on Convex; difficulty assessor and model pool built; Artificial Analysis benchmarks and per-App cost attribution TBD." />

## What it is [#what-it-is]

**Exo Router** is Exo's own smart model routing. Apps send prompts to Exo Router and never hold
provider keys. It scores how hard each task is, weighs model capability against cost, and sends
the call to the chosen model through the Convex AI Gateway. Every job is metered. In short: **an
assessor decides, and the gateway executes** ([AI routing](../../../docs/ai-routing.md)). The
design is on the [Exo Router](../platform/exo-router.mdx) page.

## Platform [#platform]

* **Convex AI Gateway** is the OpenAI-compatible egress with service-token auth. Convex holds the
  provider credentials.
* **Artificial Analysis** is named in the catalog as the source of independent intelligence,
  speed, and price benchmarks used to score models. It is **not wired into the code yet**: the
  `aiModelPool` intelligence scores (0–100) are Exo's own seed estimates in
  `convex/ai/router.ts`, "meant to be calibrated from aiRoutingLog over time". The benchmark
  import is **TBD**.
* **Arch-Router** (Katanemo Arch-Router-1.5B, open weights) is the difficulty assessor,
  self-hosted on Railway via llama.cpp from `services/arch-router/`. See
  [What is Arch-Router?](../platform/exo-router.mdx#what-is-arch-router).
* **TypeSafe** (Jev, System One) is a dormant alternate assessor: TypeSafe closed new signups
  (verified 2026-09-23).

## Boundary and replaceability [#boundary-and-replaceability]

* **Hot-swappable model pool.** `aiModelPool` is a table, so pool changes are data edits, not
  deploys.
* **Pluggable assessors.** The first configured assessor wins (`ARCH_ROUTER_URL`, then
  `TYPESAFE_API_KEY`). With neither configured, Exo Router uses pure utility and flags every
  decision `fallback: true`.
* **No silent fallback.** Gateway failures fail the job and are recorded.
* **Aggregate findings only.** Exo may receive aggregate quality, latency, token, cost, and
  reliability findings from Apps. It never receives prompts, responses, user identifiers, or
  provider keys ([Convex framework](../../../docs/convex-framework.md)).

## Cost notes [#cost-notes]

* **Assessor:** no per-call cost; CPU-only, \~1 GB RAM on Railway.
* **Routing policy:** on `staging`, `convex/ai/router.ts` computes
  `utility = intelligence^w / blendedCost`, where `w` is the assessed difficulty (0–1) and
  `blendedCost` is the 3:1 input:output blend of the pool row's per-million-token prices. Open
  PR [#115](https://github.com/Interspace-Ventures/exo/pull/115) (against `main`) changes it to
  `intelligence^w / blendedCost^(1-w)`. Easy tasks let cost dominate, and hard tasks weight
  intelligence.
* **Metering:** each pool row carries input and output cost per million tokens (USD).
  `exoAiBrokerJobs` records tokens and duration per job. On `staging` the endpoint does not yet
  compute `costMicrosUsd` and rows carry no App identifier, so per-App AI spend is **TBD**.

## Status [#status]

* **Live:** the Exo Router HTTP endpoint on the Exo Convex deployment (`POST /v1/ai/generate`),
  used by Construct. With `EXO_AI_MODEL` unset, a deterministic local composer serves requests so
  the loop works without keys.
* **Built:** `convex/ai/router.ts`, `assessors.ts`, and `chat.ts` with the routing log.
* **In progress:** sending `/v1/ai/generate` jobs through the router (PR #115), and per-App cost
  attribution
  ([roadmap: ai-gateway-jev-routing](../../../roadmap/items/ai-gateway-jev-routing.md)).
* **TBD:** Artificial Analysis benchmark import into `aiModelPool`.