Comprehensive Skills Assessment Platform

Every answer tested. Every point earned.

From algorithms to law and philosophy, every student is graded on what they know, by the same rules.

cSAP runs exams in the browser and grades them. Code and SQL meet test cases generated for every submission; written answers in any subject, from law and history to philosophy, are scored by an AI grader that must quote the answer for every point. In the exam room, cheating gets harder and easier to spot, and AI agents can run cSAP over MCP.

Request a demo See how grading works

One private deployment per institution

pair_sum.py Python

                n = int(input())
                nums = list(map(int, input().split()))
                target = int(input())
                seen = {}
                for i, x in enumerate(nums):
                    if target - x in seen:
                        print(seen[target - x], i)
                        break
                    seen[x] = i
              

Test cases

/12

Passed

Q3Why must the prosecution prove guilt?

Because the accused is presumed innocent until proven guilty, so the burden of proof lies with the prosecution.

Names the presumption of innocence+2
Puts the burden on the prosecution+1
States the standard of proof0

Score3 / 4

Test cases generated each run No quote, no point

1answer key for every programming language
3exam types: code, SQL and written
3interface languages: EN, AZ, RU
0installs: students need only a browser

A short film

Watch cSAP at work, in under a minute.

Three exam types

Code, queries and written answers. One platform grades them all.

Each exam is one type. Pick the one that tests the skill you teach, and cSAP checks every answer the same way for every student.

Programming

Students read input and print output in the language they choose at the start. You write one reference solution and one test-case generator, and they grade every language the exam offers. The generator runs again for every submission, so when it draws random input, as the templates do, each submission meets new test cases.

Python
C++
Java
C#
JavaScript
Go
Rust

Test cases generated for every submission
A time limit on every test case
A wrong answer shows the input and the first line that differs
Optional partial credit for every test case passed, best attempt kept

Step 2 of 3

SELECT name, total FROM orders
ORDER BY total DESC;

name	total
Aysel	240
Murad	180
Elena	95

Same rows, same order

SQL

Multi-step SQL tasks where every statement counts. The reference script and the student's run in separate in-memory SQLite databases, from the same setup script.

Query results compared row by row
Other statements judged by the database they leave behind
Row order matters only when the task asks for it
Optional partial credit for every correct statement

Open-ended

Short written answers in any subject, from law and history to philosophy and computing, scored by an AI grader against your rubric. A point counts only when the grader backs it with a quote that is really in the answer, and teachers can override any point.

Rubric points, added up in code
Word, character and answer-language limits
One submission per question, replaceable while it waits

How it works

From a spreadsheet to graded results

Write the exam in a spreadsheet

Each row is a problem: its description, a reference solution, a test-case generator and a time limit. A written question gets a reference answer and a rubric instead. Start from the templates and attach images as assets.

midterm.xlsx

	B	C	G	H	I
1	Title	Description	Solution	Test Cases	Max TTE
2	Pair sum	Read n numbers…	seen = {}…	import random…	1.5
3	Matrix walk	Read an n×m grid…	r, c = map(…	for _ in range…	2
4	Word count	Print each word…	from collect…	words = ["…	1
5	Bracket check	Read one line…

02
Students sit it in the browser

They sign in with the assessment name and their personal code. No accounts, nothing to install. Each student draws their own set of problems, picks a language when the exam offers several, and works in a full code editor that starts from a template for that language.
03
Grading runs in the background

Every submission goes into a queue for an isolated grader. The page updates by itself, so students can move on to the next problem while they wait. Written answers also show their place in the queue.
04
Review and share

Results come together in one table, with attendance, the average score and each student's focus-loss count. Open any student's submission history, share a read-only link or print the results.

Code	Score
1042		96
1017		88
1135		71
1096		54

Integrity

Built for exams that count

No platform can promise that nobody will cheat. cSAP makes the shortcuts hard to take and easy to see, in the grader and in the exam room.

How a programming answer is graded

1
Test cases generated every run
The generator runs again for every submission. When it draws random input, as every template does, each submission meets new test cases, and memorized outputs do not carry over.
2
Reference first, then gone
The reference solution runs first. Its output stays in the grader's memory, and its files are deleted before the student's code is written.
3
Student code runs alone
Once per test case, as an unprivileged user, under a time limit.
4
The grader decides
Pass or fail is a verdict the grader sets after comparing outputs. There is no success message a program could print to fake it.

And for written answers

Every awarded point needs a quote that really appears in the answer, the total is added up in code rather than by the model, and results are signed so a forged grade is dropped.

Exam-room controls

Focus tracking
Leaving the exam window is counted. At the limit, the student is blocked.
Copy, paste and undo off
Two switches per assessment: one blocks copy, cut and paste, the other undo and redo.
One device
Students stay on the device they started on. Teachers can reset it.
Kiosk mode
A ready-made command opens cSAP in Chrome's full-screen kiosk mode on Windows.
Personal problem draw
Every student draws their own problems from your pool, by score.
Time window
Set a start and an end. Students see the countdown.
Hidden titles
Show problems by number only.
Limited resets
Decide how many times a student may draw a new problem list.

AI grading

Written answers in any subject, graded with evidence

Open-ended questions in law, history, philosophy, literature, computing or any other subject are scored by a language model against the rubric you write. cSAP checks every quote and adds up the score, and a teacher can change any point.

Law
History
Philosophy
Literature
Languages
Computing

Your rubric, point by point
One line per point: a score and the statement it rewards. The points add up to the question's score.
No quote, no point
For every point it awards, the grader must quote the answer. A quote that is not in the answer is thrown out, and the point with it.
Arithmetic in code
The model judges each point; cSAP adds them up. The total never comes from the model.
Teachers have the last word
Override any point, or regrade one answer or a whole question after editing the rubric. Overrides record who made them and when.
Tested against tricks
Prompt injection, keyword stuffing and swapped claims are part of the test set every supported model is checked against.
Answer keys stay private
The reference answer and the rubric never reach students, and the per-point feedback is for teachers only.

The supported grader today is an open-weight model, run through a hosted provider with data collection turned off. Running a model on your own server with Ollama is built in, for models that pass the same test set.

Q2What does Descartes mean by “I think, therefore I am”?

Student answer

Descartes set out to doubt everything that could possibly be doubted, even his own senses. But doubting is itself a kind of thinking, and there must be someone who is doing the thinking, so his own existence is the one thing he cannot doubt.

Rubric

Describes his method of radical doubt
“doubt everything that could possibly be doubted”
2
Thinking proves that a thinker exists
“there must be someone who is doing the thinking”
1
It holds even if an evil demon deceives him
No evidence in the answer
1

Score 3 / 4

Multilingual

In the language your students think in

Students switch the interface and the problem text between English, Azerbaijani and Russian. Translations are optional and fall back to the original field by field, so a half-translated exam still works.

Code and test cases are written once. Only the prose is translated, so grading never changes.
Written exams can require answers in a chosen language.

English

Pair with the given sum

Read n numbers and a target. Print the positions of two numbers that add up to the target.

Input

5
3 9 1 7 4
8

Output

2 3

Azərbaycanca

Cəmi verilən ədədə bərabər olan cüt

n ədəd və bir hədəf oxuyun. Cəmi hədəfə bərabər olan iki ədədin mövqelərini çap edin.

Giriş

5
3 9 1 7 4
8

Çıxış

2 3

Русский

Пара с заданной суммой

Считайте n чисел и целевое значение. Выведите позиции двух чисел, сумма которых равна целевому значению.

Ввод

5
3 9 1 7 4
8

Вывод

2 3

Integrations and AI agents

Connected to your systems and your AI agents

Final scores flow back to the system that keeps your grades, and AI agents can run cSAP for you over MCP. Both go through the same API, with the same role checks as the admin pages.

Assessments to sync

cs101-algo-midtermSynced · 2026-03-12
cs102-sql-quiz2Not synced

GET /integration/assessments/cs101-algo-midterm/results
[{ "student_code": "1042",
   "assessment_score": 46,
   "assessment_max_score": 50, … }]

Sync scores with your student information system

Import students and subjects each semester. Your system then collects each student's final score through the integration API and marks the assessment as synced, so nobody types scores in by hand.

Students and subjects imported per semester from a spreadsheet
Each student's final score, with the maximum score and the date
Assessments filtered by synced or not yet synced

Built around one SIS so far, so connecting yours may take some adapting.

"csap": {
  "type": "http",
  "url": "https://csap.example.org/mcp",
  "headers": { "Authorization": "Bearer …" }
}

YouWho scored under 50 on cs101-algo-midterm?

assessment_results{ "assessment_name": "cs101-algo-midterm" }

Agent3 of the 42 students: 1008, 1121 and 1174.

Your AI agents, over MCP

cSAP is also a Model Context Protocol (MCP) server. Connect any agent or AI platform that reaches MCP servers over HTTP with a token, then ask in plain language: it creates assessments from a workbook, reads results and manages users and integration files.

Every tool is an API endpoint, with the same validation and role checks
Give each agent its own account, with only the roles it needs
Choose which parts of the API become tools

REST API
Assessments, users and results over HTTP, documented with OpenAPI.
One token for both
The REST API and the MCP server accept the same token, in an Authorization or x-api-key header.
Files in the tool call
Agents upload exam workbooks and images inside the tool call, the same way the admin pages do.

For teachers and admins

Everything around the exam, too

Results table
Totals, attendance, the average score and each student's passed and failed counts. Share a read-only link or print it.
Submission history
Each student's attempts, with the code or the answer and its verdict.
Exams as workbooks
Download a template, add one row per problem and upload it with its images.
Roles
Separate rights for user managers, integrations and submission analysts.
Submission log
Every passing submission is written to a log that submission analysts can read, or download as one archive.
Your own deployment
Each institution runs its own copy, with its own data and graders. Docker Compose, nothing exotic.

FAQ

Questions, answered

What kinds of exams can I run?

Three kinds. Programming exams, answered in any of the languages you offer. Multi-step SQL exams, checked against a database. Open-ended exams with short written answers in any subject, from law to philosophy, scored against your rubric. Each exam is one kind.

Which programming languages are supported?

Python, C++, Java, C#, JavaScript, Go and Rust. You choose which of them an exam offers. Each student picks one before starting, and can change it only by resetting their problem list, which you limit.

How is a programming answer checked?

You provide a reference solution and a test-case generator. For every submission, cSAP runs the generator, runs the reference to get the expected output, then runs the student's program on the same input and compares the two. Every test case has a time limit, and a generator that draws random input gives every submission new test cases.

Can students hard-code the expected output?

In programming exams it is hard to make that pay off. When the generator draws random input, every submission meets new test cases; the reference solution's files are gone before student code runs, and that code runs as an unprivileged user. In the browser, focus tracking, the copy-paste and undo switches and the one-device check make it harder still.

Can I trust AI grading?

You stay in charge of it. The grader gives a verdict on each rubric point and must quote the answer for every point it awards. Quotes that are not in the answer are rejected, and the score is added up by cSAP, not by the model. Teachers see the evidence for each point and can override it or regrade.

Do students need an account?

No. They sign in with the assessment name and the personal code you give them. A browser is all they need.

Does cSAP work with our student information system?

It can. cSAP imports students and subjects each semester, and your system collects final scores through an integration API. That API was built around one system, so connecting yours may take some adapting. The REST API and the MCP server cover anything else you want to automate.

Can an AI agent work with cSAP?

Yes. cSAP serves its admin API as an MCP server over HTTP. Give the agent its own account, with MCP access and only the roles it needs, and it can create assessments, upload workbooks, read results and manage users. Every tool call goes through the same checks as a REST API call.

Where does cSAP run, and who sees our data?

Each institution gets its own deployment with its own database, and programming and SQL answers are graded inside it. Written answers go to an open-weight model at a hosted provider, with data collection turned off. Running the model on your own server with Ollama is built in, for models that pass our test set.

How do I create an exam?

Fill in an Excel template: one row per problem or question, and a Students sheet that sets how many problems of each score every student draws. Upload it with its images, and set the time window and the exam-room controls.

Can a partly correct answer earn points?

On programming and SQL exams, if you switch on partial credit. A problem then earns its score in proportion to the test cases it passes, or for SQL the statements that are correct, and the student's best attempt counts. A fully correct answer still earns the whole score. Use it where outputs vary: an answer that prints the same thing every time earns the test cases it happens to match. Written answers always earn points one rubric line at a time.

Bring us one of your exams

Send a problem or a question you set this term, in any subject. We will set it up in cSAP and show you how it is graded: test cases, rubric and all.

Request a demo

or write to [email protected]

Every answer tested. Every point earned.

Programming

SQL

Open-ended

Write the exam in a spreadsheet

Students sit it in the browser

Grading runs in the background

Review and share

How a programming answer is graded

Test cases generated every run

Reference first, then gone

Student code runs alone

The grader decides

And for written answers

Exam-room controls

Focus tracking

Copy, paste and undo off

One device

Kiosk mode

Personal problem draw

Time window

Hidden titles

Limited resets

Your rubric, point by point

No quote, no point

Arithmetic in code

Teachers have the last word

Tested against tricks

Answer keys stay private

Pair with the given sum

Cəmi verilən ədədə bərabər olan cüt

Пара с заданной суммой

Sync scores with your student information system

Your AI agents, over MCP

REST API

One token for both

Files in the tool call

Results table

Submission history

Exams as workbooks

Roles

Submission log

Your own deployment

Bring us one of your exams