evals.sh
ProductModelsSecurityPricingDocsBlog
Sign inGet started
NewInside our computer-use agent

Evaluate software the way a senior engineer would.

Evals.sh agents use a real browser, read entire repositories and probe APIs, then score the work against your rubric with evidence for every judgement. For engineering teams and educators.

Start evaluating freeRead the docs

Free monthly credits · No card required · Never trained on your data

eval_7f3a · project-evaluation
running · step 1/9
about:blank
agent
1navigatehttps://shop-demo.app
2clickbutton "Sign in"
3typetextbox "Email" ← test@…
4typetextbox "Password" ← ••••
5clickbutton "Continue"
6clicklink "Running shoes"
7clickbutton "Add to cart"
8assertcart badge = 1
9take_screenshotcart-state.png
0%

Evaluating… collecting evidence before scoring against the rubric.

sandboxed browser · 0 actions · 0 screenshots captured
Evaluations completed
110,000+
Teams and institutions
80+
Median turnaround
32 seconds
Browser actions executed
2.8M

Platform

One platform for every kind of evaluation

Six modules share the same agent runtime, scoring engine and reporting, so results are comparable whether you're grading a cohort or reviewing a release.

Project evaluation

Point an agent at a live URL and describe what should work. It signs in, clicks through flows and verifies outcomes in a real browser.

  • Computer-use agent
  • Screenshot evidence
  • 0–100 score with rationale

Code review

Repository-wide review that reads the code as a whole, not file by file. Findings come with severity, context and a suggested diff.

  • Upload files or a folder
  • Suggested fixes
  • Single-use workspace

Frontend audit

Core Web Vitals, accessibility, SEO and security headers, with the AI ranking which fixes will move the score most.

  • LCP · CLS · TBT
  • WCAG checks
  • CSP · HSTS · framing

API testing

An agent that probes endpoints like a careful QA engineer: response shapes, error handling, auth behaviour and edge cases.

  • Schema & status checks
  • Edge-case inputs
  • Shareable report

Assignment grading

Grade submissions against your own rubric. Every criterion gets a score and a written justification students can learn from.

  • Your rubric, your weights
  • Per-criterion feedback
  • PDF / XLSX export

Bulk runs with Groups

Run any module across a whole cohort or portfolio at once. Compare results side by side and export them in one click.

  • CSV or pasted list
  • Runs in parallel
  • Sortable results table

How it works

Evidence first, then a score

Every job runs in its own sandbox. The agent loops through perceiving, planning, acting and verifying until the goal is met or proven unreachable, and only then is the evidence scored against your rubric.

Inputs flow into an isolated sandbox where the agent perceives, plans, acts and verifies; evidence is scored into a report.INPUTLive URLSource codeAPI endpointRubricISOLATED SANDBOX · PER JOBPerceivea11y tree · pixelsPlannext step vs. goalActclick · type · openVerifyassert · screenshotloop until goal metworkspace destroyed when the job endsREPORT86Score + rationaleStep-by-step logScreenshotsSuggested fixes

01

Describe the goal

A URL, repository or endpoint, plus plain-language instructions or a weighted rubric.

02

Agents do the work

They browse, read code and call APIs in an isolated environment, logging every step.

03

Review the report

A 0–100 score, written rationale, screenshots and suggested fixes, ready to share or export.

Computer use

Agents that operate software, not just read it

Static analysis can't tell you whether checkout works. Our computer-use agent drives a real browser the way a user does, adapts when the layout changes, and records what it saw at every step.

  • Selector-free. Targets elements by role, label and visible text, so it works across frameworks and redesigns.

  • Resilient. Waits for the page to settle, retries flaky interactions and recovers from unexpected dialogs.

  • Auditable. Every action, assertion and screenshot is kept in the step log attached to the report.

How the agent works
action vocabulary
sandboxed chromium
navigatesnapshotclicktypeselect_optionhoverpress_keyfile_uploadwait_fortake_screenshot

# instruction

Sign in with the test account, add any item to the cart

and confirm the cart total updates without a reload.

# agent plan

→ locate sign-in entry point

→ authenticate, confirm session

→ open a product, add to cart

→ assert total changed · capture evidence

Models

The right model for every task, picked for you

Browser agents, repository review and bulk grading stress models in different ways. Evals.sh reads what each job needs and routes it to the model best suited to it. No model menus, no guesswork.

Explore models

OpenAI

GPT family

Strong tool use and long-horizon reasoning. Our default for agentic browser sessions, where the model has to plan many steps ahead and recover from unexpected UI states.

GPT-6 LunaGPT-6 SolGPT-6 Astra

Anthropic

Claude family

Careful, well-calibrated code reasoning. Particularly good at reading large repositories and explaining why a pattern is a problem, not just that it is one.

Claude Opus 5.5Claude Fable 5.1

Moonshot AI

Kimi family

Open-weight frontier models built for long-running agents, with native vision and million-token context. A strong fit for extended browser sessions and repository-wide review.

Kimi K3

Who it's for

Built for people who review a lot of work

Educators and bootcamps

Grade an entire cohort's deployed projects and repositories in one session. Every student is held to the same rubric and receives specific, written feedback.

  • Bulk evaluation from a CSV
  • Rubric-based scoring
  • Shareable per-student reports

Engineering teams

Add an independent reviewer to releases, take-home assessments and vendor deliverables. Verify that critical flows work before customers find out they don't.

  • Release smoke-tests on live URLs
  • Repository-wide code review
  • API behaviour checks

Security & privacy

Private by design

You're trusting us with source code, student work and access to your applications. We treat that seriously: minimal retention, strict isolation, and no training on your data, ever.

Security overviewPrivacy policy

Never used for training

Your code, URLs, rubrics and results are never used to train models, by us or by our model providers.

Ephemeral workspaces

Code you submit is placed in a single-use workspace that is deleted as soon as the job finishes.

Isolated execution

Every browser session runs in its own sandboxed container. Nothing is shared between jobs or customers.

Encrypted end to end

TLS for all traffic in transit and encryption at rest on enterprise-grade cloud infrastructure.

Least-privilege access

Every request is authenticated and scoped to your account. Report share links are only created when you ask for one.

You own your data

You keep ownership of everything you submit. Ask us to delete a report or your account at any time. We never sell customer data.

Pricing

Simple, credit-based pricing

Every feature is available on every plan. Plans give you discounted credits. Top up any time at $1 = 5 credits.

Free

Every feature, to get started

$0/month

  • 10 Code reviews/mo
  • 10 Frontend audits/mo
  • 10 Project evaluations/mo
  • Basic models
Start free
Popular

Starter

For individual instructors and small teams

$20/month

  • 200 Code reviews/mo
  • 200 Frontend audits/mo
  • 200 Project evaluations/mo
  • Intelligent model picking
Contact sales

Pro

For departments and engineering orgs

$100/month

  • 2,000 Code reviews/mo
  • 2,000 Frontend audits/mo
  • 2,000 Project evaluations/mo
  • Intelligent model picking
Contact sales

FAQ

Do you use my code or my students' work to train AI models?
No. Customer data is never used to train models, by us or by the model providers we work with. Requests are processed under enterprise terms that exclude training.
What happens to my code after a code review?
Your code is placed in a workspace created for that single job, and the workspace is deleted when the job completes. The review report is kept in your account so you can revisit it.
Can the agent test apps behind a login?
Yes. Include test credentials in your instructions and the agent will sign in like a user. We recommend dedicated test accounts rather than real ones.
How consistent are the scores?
The agent gathers evidence first, then scores it against your rubric or instructions. Clear, observable criteria produce the most repeatable results. Our rubric guide covers how to write them.

From the blog

All posts →

Engineering · October 23, 2026

Inside our computer-use agent: evaluating apps through a real browser

Security · October 23, 2026

What happens to your code when you run an evaluation

Product · October 23, 2026

Matching the model to the evaluation

Your next review takes minutes, not a weekend.

Create an account and run your first evaluation on the free plan.

Get started freeTalk to us
evals.sh

Autonomous evaluation for software teams and educators. Agents that review code, drive real browsers and grade work against your rubric.

Product

  • Project evaluation
  • Code review
  • Frontend audit
  • API testing
  • Assignment grading

Platform

  • Models
  • Security
  • Pricing
  • Documentation

Resources

  • Blog
  • Guides
  • Contact

Legal

  • Privacy policy
  • Terms of service
  • Data handling

© 2026 Evals.sh. All rights reserved.

Customer data is never used to train AI models.