What Gate does
You give Gate a piece of AI-written work — code, or a plan for work not yet started. Gate checks it and returns a verdict, a score, a list of findings and suggestions for fixing them. Two kinds of checking are involved, and it helps to keep them apart:
- Deterministic checks. Fixed checks for problems such as placeholder or mock implementations presented as finished, calls to APIs that do not exist, and unsupported claims of completion. Any one of these can block a result on its own.
- Model-assisted engineering review. A language model reviews what survives for performance, reliability, security and maintainability risks. This review is informed judgement, and it can miss things or raise findings you disagree with.
Connect Gate
Gate connects to AI tools using MCP (Model Context Protocol), a standard way for AI assistants to call external tools. One-click install links for Cursor, VS Code, LM Studio and Goose are on the product page. For any other MCP client, add this to its MCP settings:
{
"mcpServers": {
"verificate": {
"url": "https://mcp.verificate.ai/mcp",
"transport": "http"
}
}
}In Claude Code, one command does the same:
claude mcp add --transport http verificate https://mcp.verificate.ai/mcpThe first 25 checks on a machine need no account and no token. After that, create an account for a 30-day trial (no card required) and add your token to the same configuration:
{
"mcpServers": {
"verificate": {
"url": "https://mcp.verificate.ai/mcp",
"transport": "http",
"headers": { "Authorization": "Bearer <YOUR_TOKEN>" }
}
}
}The four tools
| Tool | Use it when |
|---|---|
validate_ai_output | An assistant has produced code or another output and you want a verdict before accepting it. |
validate_plan | An assistant has proposed a plan and you want it checked for feasibility and grounding before work begins. |
analyze_code | You want a deeper review of performance, scalability, reliability and technical debt. The review is informed by the ISO/IEC 25010:2023 quality model. |
generate_code | You want code generated and checked in one step, with the checking applied before anything is returned. |
Your AI tool lists each tool’s exact inputs once Gate is connected; that live listing is the reference to rely on.
Reading a verdict
Swipe sideways to see all 3 columns.
| Verdict | What it means | What to do |
|---|---|---|
| Approved | No blocking problems were found. | Proceed with your normal review. Approval is not proof the work is free of defects. |
| Approved with warnings | Nothing blocked the result, but there are findings worth attention. | Read the findings. Decide which to fix now and which to record. |
| Rejected | A blocking problem was found. | Do not merge or present the work as finished. Fix the findings and check again. |
Alongside the verdict you receive a score from 0 to 100, a list of issues — each with a severity and a category — and suggestions. Read the issues rather than the score: a single critical issue matters more than a number.
Acting on findings
- Start with blocking and high-severity findings.
- Fix the cause, not the wording. Removing the word “placeholder” from a placeholder does not make it real.
- Run the check again on the changed work. A fix counts when the same check passes.
- If you believe a finding is wrong, record why. Model-assisted findings are advice; you remain responsible for the decision.
A useful standing instruction for an AI coding assistant is: “Validate your output with Verificate Gate before presenting it as finished, and treat a rejected verdict as authoritative.”
Pull requests and pipelines
Gate can run as a required check on pull requests, so a rejected change cannot be merged until it is fixed. Setup instructions are maintained with each integration:
- GitHub Action (opens a new tab) — a required check on every pull request.
- LangChain integration (opens a new tab) — check agent and chain outputs before they reach downstream systems.
- More integrations are listed on the product page.
If you commission software
You do not need to read code to use the results. Ask your supplier to run Gate on what they deliver and share the verdict and findings, or run a check on a repository yourself. Findings are grouped in business terms:
- Security — could this expose data or be misused?
- Reliability — what happens when something fails?
- Performance — will it cope with real usage?
- Maintainability — can someone else change it safely later?
Use the findings as questions for your supplier before sign-off, not as an automatic pass or fail on the whole engagement.
What a verdict is not
- Not a certification. Gate does not certify software as safe, secure or compliant with any standard.
- Not a conformance claim. Gate is working toward an ISO/IEC 5055-aligned gate; it does not currently claim ISO/IEC 5055 conformance or complete coverage of its measures.
- Not a replacement for testing, security review or human code review.
- Not exhaustive. An approved result means no blocking problem was found in what was submitted — not that none exists.
What you send
Gate reviews the content you submit to it. Submit only work you are permitted to share, and leave out secrets, credentials and personal data. For how information is handled, see the privacy policy and security information, or ask us about your requirements.
Plans and trial
There are 25 free checks per machine with no account, then a 30-day trial with no card required, then a monthly subscription. Current prices and offers are on the product page. Volume and academic pricing is available on request.