Agents make promises. We check which ones they keep. Suncly is an independent lab that tests whether your MCP server or agent does what it declares, and signs the evidence.

Research · Approach

Our approach to agent verification

Agents now act for people, and they describe themselves to every system they join. We explain why we test agents against their own declarations, how we decide whose failure a failure is, and why every gap belongs in the report.

Read our approach
Product · Available now

Suncly Verification

A one-off test of an MCP server against its tool list, or an A2A agent against its Agent Card. See the price before anything runs; pay when the signed report arrives.

Sample reportPricing
Start a test
Product · Early access

Suncly Benchmark

Your users’ tasks, run end to end in ChatGPT, Claude and Copilot every month. How often each one worked, why it failed, and what changed since last month.

Sample benchmarkHow it works
Join early access
News · 11 Nov 2026

Agent incidents, analysed

We follow attacks and accidents involving AI agents and MCP servers, and say what an evaluation would have shown. The latest: an MCP gateway that could be steered at the network it guards.

An MCP gateway steered at its own networkCoding agents install unpatched MCP servers
All news
Lab · Updated 4 Nov 2026

What we can test today

Read and create tools are tested in full. Update and delete run only on objects we created. Send, invite and injection tests are listed as not yet, in every report.

Coverage by kind of toolLab notes
Visit the Lab

Two ways to get the evidence

Verification checks one server or agent against what it declares, once. Benchmark, in early access, answers whether your agent works for your users, every month. Either way, a person approves every plan before anything touches your system.

Open the console
VerificationOne-off · Available nowHow it works
01Paste a URLYour MCP server or Agent Card. We read every tool and skill it declares.
02See your priceWorked out from your tool list, before anything runs. Read tools cost least.
03Approve the planEvery case is listed. Remove any you don’t want. Your card is held, not charged.
04Get the reportSigned, with every request and response. You’re charged when it arrives.
BenchmarkMonthly · Early accessHow it works
01Write your scenariosYour users’ tasks, in their words. We can draft them from your support tickets.
02Pick the clientsChatGPT, Claude and Copilot. Each scenario runs twenty times in each one.
03Set the standardWhat counts as done, decided once, before anything runs.
04Read the monthly reportA pass rate with its N, failures by cause, and what changed.
// Inside the Lab

Every square is a promise, checked.

One square per run of a user task, one row per scenario, grouped by client. Green got the task done, amber couldn’t be decided and never counts as a pass, and terracotta failed, with its cause tagged: prompt, skill, tool or client UI. This is what a Suncly run looks like while it happens, and exactly what the signed report shows when it’s done.

Visit the Lab
A paper half-sun beside a ruler, in soft daylight
Illustrative run · 6 scenarios × 3 clients · the first 8 of 20 runsDone   Couldn’t decide   Failed, cause tagged   Not run yet
Learn more
Our approachWhy agents should be judged against their own declarations.
The LabHow a test runs today, and what we can and can’t yet test.
Our commitmentsThe rules we hold ourselves to in every test and report.
Agent incidentsRecent attacks and accidents, and what an evaluation would have shown.