Checkcraft

Desktop runner for API, Kibana-log and database checks. The agent drafts checks from requirements; Checkcraft stores and runs them the same way every time and reports expected vs actual. Secrets are never sent to the agent. Windows, free tier.

Documentation

Checkcraft is a desktop app for checking APIs and logs without code, with a free edition. Download, no sign-up →

An AI agent writes tests quickly, but it also "fixes" them by bending the expectation to the response. Checkcraft splits the roles: the agent drafts the checks, the app runs them the same way every time, and you accept the report.

Get help with your first test →

Why an agent needs a separate executor

When the same agent writes and runs a test, it has a shortcut to green: rewrite the expectation to match the actual response. The test then passes and proves nothing. With Checkcraft the agent only drafts checks and reads the report. The app runs them from saved rules: the same data gives the same result, and every change the agent makes is visible in the window.

How to connect

Download Checkcraft for Windows 1.0.185 or later; the free version works. Open Settings → "AI agents" and turn on "Allow AI agents to work with Checkcraft". The same tab shows the connect command with the full path to checkcraft-mcp.exe: copy it into a terminal. In Cursor and other MCP clients, use this file as the server command. The agent works while the app is open.

# Claude Code: copy the path from Settings → "AI agents" → "Copy"
claude mcp add checkcraft -- "C:\Checkcraft\checkcraft-mcp.exe"

# Cursor and other MCP clients: the server command is the same file
{ "mcpServers": { "checkcraft": { "command": "C:\\Checkcraft\\checkcraft-mcp.exe" } } }

What the agent can do

Look at existing checks; create a check from a description or a cURL command; build a test case from Checkcraft steps — status, response fields, calculations, arrays, Kibana log records by traceId; run a check and get a report where each failed rule shows the expected value; search log records by traceId; undo its own last change. In Pro the agent adds a database row check (SELECT only) and runs a Pro scenario: the request, the test cases and the database comparison in one run.

An example request to the agent

Requests with expectations taken from requirements work best, rather than "check that everything is fine". The agent creates a check and a test case, runs it and comes back with the differences. The expected values in the rules stay as in the requirements, so a failure stays visible instead of being papered over.

Requirements for POST /orders: 201 response, status = NEW,
total equals the sum of price × qty over items, no ERROR records in the
service logs for the traceId. Create a check and a test case in Checkcraft
and run it. If something does not match, show expected and actual.

Security

Access is off by default. The app accepts agent calls only from this computer and only with a key stored in your user folder; browser requests are rejected. Passwords, tokens and database connection strings are not passed to the agent: they are masked in responses, and the agent cannot create a request with a plain secret — only with a placeholder such as {{TOKEN}} whose value you enter. Each agent change is visible in the window, can be undone from the notification, and the settings can require confirmation for every change. The agent and its model remain yours: whatever you write to it goes wherever you configured it to go.