Comprehensive Skills Assessment Platform
cSAP runs programming, SQL and open-ended exams in the browser and grades them for you: code in many languages against test cases generated for every submission, written answers by an AI grader against your rubric, with the evidence behind every point. It connects to your student information system, and AI agents can run it over MCP.
One private deployment per institution
Q3Why does an index make this query faster?
An index lets the database find matching rows without reading the whole table, because it keeps the keys sorted in a B-tree.
Score3 / 4
Three exam types
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.
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.
Multi-step tasks where every statement counts. The reference script and the student's run in separate in-memory SQLite databases, from the same setup script.
Short written answers, 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.
How it works
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.
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.
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.
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.
Integrity
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.
The generator runs again for every submission. When it draws random input, as the templates do, memorized outputs do not carry over.
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.
Once per test case, as an unprivileged user, under a time limit.
Pass or fail is a verdict the grader sets after comparing outputs. There is no success message a program could print to fake it.
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.
Leaving the exam window is counted. At the limit, the student is blocked.
Two switches per assessment: one blocks copy, cut and paste, the other undo and redo.
Students stay on the device they started on. Teachers can reset it.
A ready-made command opens cSAP in Chrome's full-screen kiosk mode on Windows.
Every student draws their own problems from your pool, by score.
Set a start and an end. Students see the countdown.
Show problems by number only.
Decide how many times a student may draw a new problem list.
AI grading
Open-ended questions 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.
One line per point: a score and the statement it rewards. The points add up to the question's score.
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.
The model judges each point; cSAP adds them up. The total never comes from the model.
Override any point, or regrade one answer or a whole question after editing the rubric. Overrides record who made them and when.
Prompt injection, keyword stuffing and swapped claims are part of the test set every supported model is checked against.
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.
Q2Why can deep recursion cause a stack overflow?
Student answer
Each recursive call adds a new frame to the call stack, and the frames are only removed when the calls return. If the recursion goes too deep, or there is no base case, so it never stops, the stack runs out of space.
Rubric
Each call adds a stack frame
“adds a new frame to the call stack”
A missing or unreachable base case
“no base case, so it never stops”
Names a fix: iteration or a smaller depth
No evidence in the answer
Multilingual
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.
The same problem in three languages
English
Read n numbers and a target. Print the positions of two numbers that add up to the target.
5 3 9 1 7 4 8
2 3
Azərbaycanca
n ədəd və bir hədəf oxuyun. Cəmi hədəfə bərabər olan iki ədədin mövqelərini çap edin.
5 3 9 1 7 4 8
2 3
Русский
Считайте n чисел и целевое значение. Выведите позиции двух чисел, сумма которых равна целевому значению.
5 3 9 1 7 4 8
2 3
Integrations and 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 · 2026-03-12cs102-sql-quiz2Not syncedGET /integration/assessments/cs101-algo-midterm/results [{ "student_code": "1042", "assessment_score": 46, "assessment_max_score": 50, … }]
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.
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.
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.
Assessments, users and results over HTTP, documented with OpenAPI.
The REST API and the MCP server accept the same token, in an Authorization or x-api-key header.
Agents upload exam workbooks and images inside the tool call, the same way the admin pages do.
For teachers and admins
Totals, attendance, the average score and each student's passed and failed counts. Share a read-only link or print it.
Each student's attempts, with the code or the answer and its verdict.
Download a template, add one row per problem and upload it with its images.
Separate rights for user managers, integrations and submission analysts.
Every passing submission is written to a log that submission analysts can read, or download as one archive.
Each institution runs its own copy, with its own data and graders. Docker Compose, nothing exotic.
FAQ
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, scored against your rubric. Each exam is one kind.
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.
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 case has a time limit, and a generator that draws random input gives every submission new cases.
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.
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.
No. They sign in with the assessment name and the personal code you give them. A browser is all they need.
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.
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.
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.
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.
Send a problem or a question you set this term. We will set it up in cSAP and show you how it is graded: test cases, rubric and all.
or write to sales@csap.app