AI writes the code. The spec decides what ships.
Coding agents write most of the code now, from a spec that lives in a chat window, and then mark their own homework. Test Maze holds the user stories and acceptance tests, keeps the agent inside them, has a separate verifier grade the work with fixed rules, and only opens the release gate on the results.
Why we built it
Coding agents are fast, and they are convincing. Ask one whether the change is done and it will tell you yes, usually with a tidy summary of tests it believes it ran. There is no independent record: the requirements, the tests, the results and the verdict all come from the same place that wrote the code. Three things go wrong, every time — the spec drifts from one session to the next, code lands that nobody asked for, and flaky work hides behind a confident paragraph.
We split the job. People own the story and the acceptance criteria and approve what agents propose. Your agent keeps the source tree and builds inside the declared story. A separate verifier runs the tests, and Test Maze stores each result against its git sha and grades the change deterministically. When the release gate opens, you can point at the story, the commits, the run and the empty defect list that opened it.
How the loop works
Everything happens over the Model Context Protocol (MCP), so it works with Claude Code, Cursor, Cline, Gemini CLI, Codex CLI and any other MCP client. You talk to your agent in plain English; the agent calls the tools.
- Spec first.
feature.implementturns a feature title into user stories and acceptance criteria;feature.verifymakes them runnable test cases. A person reviews them and ticks Definition of Ready. - Every change traced. With spec-based development on,
spec.start_workdeclares the story for the branch; until then the Claude Code hook blocks edits. Every commit carries aTestmaze-Ref; CI re-checks the range. - Verified independently. The
testmaze-verifiersubagent runs the cases and records them withtestrun.record_results, pinned to the git sha.pdlc.verifyreturns the deterministic verdict and a next step. - Shipped on evidence. Failures become defects (
defect.create, with a regression case and a GitLab issue).release.readinesschecks the runs and open defects;release.shipis refused while the gate fails.
Around that spine sit exploratory testing with the agent's own browser, coverage gap analysis, frozen regression baselines, acceptance-criterion waivers, the traceability matrix and code-quality checks — 70 tools in all.
What we hold to
Agents propose, people approve
Everything an agent creates starts as Pending. Reviewing a story or case, ticking Definition of Ready and Done, and shipping a release past a failing gate are human-only actions in the web app; there is no MCP tool for any of them.
The verdict is computed, never guessed
pdlc.verify grades a change with fixed rules over the recorded test results: pass when at least one case passed and none failed, fail when a case failed or a frozen regression baseline is missing or failing. The same results always give the same verdict. AI may add a hint about what to fix — it never decides.
We never read your repository
Test Maze does not clone, scan or index your code. What reaches it: the product summary your agent writes, stories, test cases and results, git identifiers, optional screenshots, any code snippet an agent chooses to paste into a code-quality check, and — only when spec-based development is on — commit subject lines and Testmaze-Ref trailers. Never diffs, never files.
Every result is tied to a commit
Each verifier test run records the git sha, the branch and whether the working tree was clean. A green run means a specific commit was green — not "it worked on someone’s machine at some point".
Your agent stays in charge
The coding agent owns the source tree and runs the tests. Test Maze owns the test artefacts and the grading. Neither side has to trust the other’s summary of what happened.
Who it is for
Everyone who ships with AI and still has to answer "how do you know it works, and was it what we asked for?" — vibe coders who want the speed without the doubt, developers who want to write the story once, QA engineers and product owners who keep authority over the acceptance criteria, engineering leads who want to govern agents without slowing them, and regulated teams who need the chain from requirement to release on one page.
Who builds it
Test Maze is built by Ceegees, a software engineering company that has spent years building and testing products for other teams. We wanted the tool for our own agent-assisted work first.
Questions, feedback or a pilot for your team? Talk to us — or read the FAQs and pricing.
Ready to give your agent a verifier?
Create a free workspace, generate an MCP token and connect your coding agent in under five minutes.