The perception layer for AI

Give your AI agents
the five senses.

A headless agent can reason — but it's blind, deaf, and numb to what its software actually did. Pentasor is the sensory nervous system that closes the gap: senses transduce a target into one common signal, and a predictive cortex binds them into a single explainable verdict.

Plug in via MCP, API, or SDK — no matter which AI you build on.

pentasor perceive — live
$pentasor perceive
✓ OK https://example.com
sensestouch, electroception
surprise███░░░░░░░ 0.314 confidence 0.66
why:
· bound 2 signals into 1 temporal group
↯ electroception@0s security.missing=csp, hsts, x-frame-options … (5 absent)
· fused surprise 0.314 vs threshold 0.5 → OK
↑ type any URL and hit run — this is live.
The sensorium

Five human senses — and beyond

Penta means five. The five human senses are the branded foundation — what make the agent human. But an AI isn't bound by biology, so the registry is open: any new sense plugs in through the same contract. Five is the foundation, not the ceiling.

👁
visual

Sight

what rendered

ffmpeg frames from media + live DOM-state probes (Playwright)

👂
audio

Hearing

what was said or played — and exactly when

Whisper → one signal per spoken segment, timed for cross-modal binding

haptic

Touch

how it responds when poked

active probing: HTTP latency / status, payload weight, interaction feedback

👃
metric

Smell

what's subtly off

log scent-sniffing + robust z-score on metrics — anomaly = prediction error

👅
quality

Taste

whether it is actually any good

frame luminance/contrast quality now; vision-model rubric when keyed

sense+ · superhuman
🫀
somatic

Interoception+

the agent's own body

process heap / RSS / CPU / load — perception turned inward

temporal

Chronoception+

a whole timeline at once

event-trace span / gaps / order / jitter — superhuman, non-serial time

field

Electroception+

invisible signal fields

HTTP headers, redirects, security posture, env shape

A nervous system, not a toolbox

How perception happens

No single sense returns a conclusion — just as no retina decides what a face is. Senses transduce into a common signal; the brain binds them into perception. Pentasor mirrors the two layers, grounded in the neuroscience of how brains actually perceive.

01the organs

Transduce

Each sense converts its raw stimulus — pixels, audio, DOM, metrics, anything — into a common Signal: extracted features, never raw bytes (the retina ships edges, not light). One modality-agnostic currency.

Sense.transduce(stimulus) → Signal[]
02the cortex

Bind & predict

A predictive cortex holds an expectation and measures surprise. It binds signals across senses in a temporal window — what was seen at t against what was heard at t — and fuses them weighted by reliability (∝ 1/variance).

Cortex.bind(Signal[]) → Percept
03the verdict

Explain

Out comes one Percept: a pass/fail with a surprise score, the senses that fired, the evidence, and the why. Every verdict shows its work — so a wrong reading is visible the moment you read it.

predict → bind → fuse → surprise

adding the Nth sense = sensorium.register(yourSense) — the cortex integrates it, nothing else changes.

Plug in any AI

One perception layer. Any agent, any builder.

Whether you write code or you don't, Pentasor connects to the AI you already use. The same queryable cortex — perceive X, return the evidence — is exposed through every surface, so new senses light up everywhere at once.

For buildersMCP · API · SDK · CLI

Drop it into your agent

Native MCP so any client — Claude Code, Cursor, Cline — invokes it autonomously. A REST endpoint for any platform. A typed SDK for embedding.

MCP — one entry in your client config
"pentasor": {
  "command": "npx",
  "args": ["pentasor-mcp"]
}
API — perceive from anywhere
curl pentasor.com/api/perceive \
  -d '{ "target": "https://app.example.com",
        "intent": "checkout renders and responds" }'
SDK — embed it
import { createDefaultSensorium } from 'pentasor'
const p = await sensorium.perceive(target)
For everyoneno code required

Or just connect and look

Paste a link or drop a file and watch Pentasor perceive it — in plain language, with the evidence shown. Connect your assistant in one click; no install, no setup.

🔗
Paste a URL or drop a file
a page, a video, a build — anything your AI made
Connect your AI in one click
Claude, ChatGPT, Cursor — Pentasor becomes a sense it can use
🧾
Read the verdict in plain words
“the checkout looks right, but it’s slow and missing two security headers” — with proof
The app where anyone perceives
app.pentasor.com — coming with early access
Grounded in neuroscience

Not a metaphor.
The real architecture of perception.

Biology already solved “many senses → one understanding.” Pentasor builds that formula: transduce, encode to a common code, predict and measure surprise, fuse by reliability, bind in time, and act where uncertainty is highest.

Perception = prediction error

The cortex predicts what the target should do and measures the surprise. Verification is predictive coding (Friston).

Fuse by reliability ∝ 1/variance

Senses are combined by inverse-variance weighting — statistically optimal multisensory integration (Ernst & Banks).

Bind in a temporal window

Co-occurring signals are one event: what's seen at t binds to what's heard at t (the unity assumption / McGurk).

Any stream can become a sense

Sensory substitution proves it — a brain learns to see through the tongue. New senses just transduce into the same currency.

For AI platforms

The perception layer, ready to go native

Today's computer-use perceives interfaces by dumping screenshots into a vision loop — slow, costly, brittle. Pentasor transduces the interface semantically into a compact signal: extract features, don't ship pixels. Fast, cheap, explainable, and cross-device — one transducer per platform, one common cortex.

Path A · Embed

Native support inside your stack

Wire Pentasor in as the perception primitive behind your agent's eyes, ears, and instincts. Your models reason; Pentasor sees, hears, and senses what actually happened — and returns an explainable signal, not a screenshot.

  • A perception API your agents call
  • Cross-device: web, desktop, mobile
  • Explainable signals your model can trust
Path B · Offer

Available in your ecosystem

Ship Pentasor to your users as a first-class MCP server / connector in your marketplace. They get senses on day one; you get a perception capability without building the sensory stack yourself.

  • Listed MCP / connector
  • Zero sensory infrastructure for you
  • Your users verify what their agents build
Building agents at Anthropic, OpenAI, or beyond?

We're proving it on real verification work first. Let's talk about native perception.

Start a conversation →

Let your agents perceive what they ship.

Request early access. Tell us how you'd plug in — we're onboarding builders and teams first.