SpiderUI for agents.

UI for agents

Spider sits between your agent and the browser. It shows the agent a compact view of each page, turns finished tasks into recipes it can replay, and checks every result. On our test portal, agents read 92% fewer tokens.

Get early accessHow it works

MCP · REST API · TypeScript SDK · local or hosted Chromium

Why NOW

Consumer AI is becoming agents. And agents live in the browser.

The next wave of AI isn’t a chat window. It’s agents that do things for people: book, buy, pay, cancel, file. Almost all of that happens on websites that will never ship an API.

At consumer scale, every page an agent re-reads, often 56–80K tokens of HTML, is paid for at model prices, millions of times a day. Cheap, reliable browsing stops being a feature and becomes infrastructure.

  1. 01 · Consumer AI goes agentic

    900M

    people use ChatGPT every week. OpenAI’s Atlas and Perplexity’s Comet now give them agents that browse and act on websites.

  2. 02 · Agents finally work

    61%

    of 300 live-website tasks completed by OpenAI’s Operator in 2025, up from about 30% for earlier agents.

  3. 03 · The hard part moved

    58%

    of Operator’s failures were filtering and sorting mistakes, not clicking the wrong thing. The gap is reliable actions and checked results.

  4. 04 · The rules aren’t set

    97%

    of llms.txt files, the “agent-friendly” files sites publish, got zero requests in a month in one sample. Agents have to work with sites as they are.

Sources: OpenAI usage as reported in Feb 2026; Online-Mind2Web (2025); W3C WebMCP draft and llms.txt usage data, compiled in our 2026 research notes.

Measured

92% fewer tokens read. Same task, same steps.

Two invoice tasks, run on five versions of our test portal: renamed buttons, moved controls, a pop-up banner, and slow loading. Every agent took the same actions. Only what it read back from the browser changed.

tokens read from the browser · 10 runs

  • Normal agent (full snapshot)122k
  • Agent reading HTML127k
  • Agent + Spider9.4k
  • Recipe replay0

cost with prompt caching · 10 runs

  • Normal agent (full snapshot)$1.42
  • Agent reading HTML$1.45
  • Agent + Spider$0.66
  • Recipe replay$0.00
92%
fewer tokens read
13.1× less for the agent to read back from the browser.
2.4×
cheaper per run
2.1× with prompt caching on, priced at Claude Opus 5.5.
$0
for a recipe replay
A warm run costs the customer’s agent no tokens at all.
20/20
runs verified
10 first runs and 10 recipe replays, all passing their checks.

Measured on our mock portal, whose pages are small; real portals are larger, so the gap in tokens should grow. Both agents took identical steps rather than a live model choosing them, and tokens are estimated as characters ÷ 4.

Where Spider fits

Agents don’t need another browser. They need a better way to use one.

Caching and replay already exist elsewhere. Spider’s difference is the combination: compact views, recipes saved only from verified runs, repairs saved as new versions, rule-based checks, and a cost record for every run, behind one MCP or REST interface.

  • Cloud browsers

    • Browserbase
    • Kernel

    Host Chrome sessions for agents at scale.

    Works on top of them: point Spider at a hosted browser and it adds compact views, recipes, and checks.

  • Agent frameworks

    • Browser Use
    • Stagehand

    Ask a model what to do in Chrome, and cache or replay what worked.

    Saves recipes only from runs that pass their checks, repairs them into new versions, and labels every result.

  • New engines

    • Lightpanda

    Fast headless engine without real visual layout.

    Runs real Chromium today, and is building its own engine with native hooks for agent state.

  • Request replay

    • Unbrowse

    Turns a site’s network traffic into callable tools.

    Runs the real page, so scripts, forms, and logins behave as they do for people.

Competitor capabilities as documented by each vendor, Nov 2026. Not head-to-head benchmarks.

01 · Read

Show agents only what matters.

Your agent connects over MCP or REST. Spider drives a real Chromium browser, locally or hosted on Browserbase or Kernel, and hands back a compact view of each page plus a short report of what changed after every action.

  • Stable IDs for every button, field, and table
  • 13× fewer tokens read than full page snapshots
  • Risky actions wait for a person to approve
spider · chromiummock vendor portal

01 sign insaved login

Spider signs in to the mock vendor portal, filters unpaid invoices from October 2026, reads seven rows, downloads each PDF, checks that each file holds the right invoice number, and saves the run as a recipe: USD 15,579.00 and EUR 3,042.00.

02 · Replay

Solve it once. Replay it next time.

When a run passes its checks, Spider saves it as a recipe with parameters, row loops, and a fingerprint for every element. Replays re-find each element and call a model only at judgment points.

  • Moved or renamed controls are re-found and saved as a new version
  • If a step can't be re-found, it is reported, never guessed
  • A recipe replay costs your agent zero tokens
recipe · unpaid_invoices(month)v2
  1. gotoportal /invoices
  2. actselect [Status] "Unpaid"
  3. actfill [From] {month.start}
  4. forEachrows like INV-*next page until gone
  5. actclick [Download PDF]
  6. checkfile contains {item.key}
  7. judge"is this a credit memo?"only model call
  8. checkcount matches the site
example repair v1 → v2: [Download PDF] renamed, re-found by role + row

03 · Verify

Checked, not claimed.

Every run ends with rule-based checks: the file downloaded and holds the right invoice number, and the count matches what the site shows. Unconfirmed work is never reported as success.

  • Verified complete, verified partial, failed, or unverified
  • Totals summed per currency from verified rows only
  • Every step, model call, and repair logged with its cost
run_recipe · example resultverified_complete
items
7 of 7 ok · INV-1016 … INV-1031
files
7 PDFs, invoice id found in each
totals
USD 15,579.00 · EUR 3,042.00
repairs
1 · refind, confidence 0.94
unresolved
none
cost
agentTokens 0 · 1 model call (judgment)
verified_completeverified_partialfailedunverified

Next: our own engine

A browser built for agents, from the engine up.

Today Spider runs on Chromium. In parallel we’re building its own engine, a modified Servo, so agents can see and check page state directly instead of reconstructing it. A proposal never grants itself permission, and a click never counts as success.

  1. 01 · Decide

    AI planner

    in progress

    model_worker

    Reads page records and proposes a typed program.

  2. 02 · Check

    Program compiler

    working

    agent-executor

    Checks types, effects, and permissions before anything runs.

  3. 03 · Act

    Runtime + journal

    working

    agent-runtime-service

    Runs permitted steps and journals side effects before they are sent.

    Spider engine

    working

    patched Servo

    Real JavaScript, layout, cookies, and downloads.

  4. 04 · Prove

    State store

    working

    agent-state

    Versioned records of what the page showed.

    Independent verifier

    working

    portal oracle

    Checks real files against the expected invoices and writes the receipt.

working: demonstrated on the team’s local test portal in progress: built and tested, not yet run with a real model

6/6
full invoice runs passed
Sign in, search, page through, export, download, verify.
15/15
exported files verified
Checked by a separate verifier, not by the actor.
20/20
download checks
Normal downloads plus expired login, truncated file, unsafe filename, wrong account.
16/16
session checks
Wrong account, expired session, cookie races, across two accounts.

Engine prototype results, on its own local test portal, with programs written by the team and zero model calls. Next milestone: the AI writes the program, reuses it, and repairs it when the page changes.

Bring us a task your agent runs every day.

We’re working with a small group of teams running browser agents in production, and measuring Spider against their current setup on real tasks.

Email the foundersHow it works