Docs

MCP Tester is a hosted service that tests an MCP server the way the AI assistants that will use it do. You give it a server's address; it connects, reads every tool the server declares, and sends a crew of AI agents to use them. You get back one report: what works, what breaks, and how to fix it.

It needs nothing from you except the address. Your server doesn't have to be listed anywhere, and there is no SDK to install.

Paste your server's address, look at what it declares, then start a run. None of the three steps needs an account.

1. Paste the server's address on the home page. If the server asks for a login, approve access in the window that opens, or paste an API key.
2. Look at what it declares. The server's page lists every tool, with its arguments and schemas.
3. Set up a run. Pick the agents you want, then press Run agents →. Each agent's progress shows as it works, and the report fills in when it's done.

Sign in with Google at any point to keep your work, add teammates and see past runs.

Any remote MCP server that speaks Streamable HTTP, the transport in the current MCP specification. MCP Tester doesn't run local (stdio) servers, and it doesn't support the older HTTP+SSE transport.

• The address must start with https://.
• Addresses inside a private network, and addresses with a username or password in them, are refused.
• Servers on MCP protocol versions from 2025-03-26 onwards are supported.

If a connection fails, the message tells you where it stopped:

You see	What it means
Couldn't connect. Is this an MCP server?	Nothing at that address answered as an MCP server. Check the address and its path.
This server takes a key, not OAuth. Paste it below.	The server wants a login but has no OAuth sign-in. Paste an API key instead.
Its OAuth refused to register this client.	The server only accepts clients registered in advance. We can't sign in to it yet.
Its OAuth metadata names a different URL.	The server's sign-in is set up for another address, so a login here wouldn't work.
Connected, but the tool list never finished.	The server said hello but didn't finish listing its tools.
The server speaks an MCP version we don't support.	Its protocol version is outside the supported range.

With OAuth when the server offers it, or with an API key you paste when it doesn't.

• OAuth. We follow the MCP authorization spec (OAuth 2.1 with PKCE). We identify ourselves with a client metadata document where the server supports one, and otherwise register a client automatically. The server's own consent screen opens in a new window; approve it there.
• API key. For servers without OAuth, paste the header name and value. We send them on every request to that server, and to nothing else.

A login belongs to the person who gave it. Teammates who open the same server sign in with their own login, and a run always uses the login of the person who started it. Logins are encrypted when stored.

To remove your login, press Disconnect on the server's page. That deletes it from MCP Tester; to revoke it on the server's side too, use the server's own settings. If your login might expire before a run finishes, the run page warns you before you start.

Everything the server declares, read live every time you open the page.

• The server's name, icon, protocol version and the kind of login it uses.
• The instructions it gives assistants, if it gives any.
• Every tool: its name, description and arguments, a rough count of the tokens its definition costs an assistant, and its full input and output schemas.

Refresh reads it all again, for example after you deploy a change. Resources and prompts aren't listed yet.

The agents you pick call your server's real tools, then write up what they found.

1. Press Set up a run → on the server's page. Your setup is saved as you go, so you can leave and come back to it.
2. Pick the agents. If you pick happy-path, also pick up to three assistant models for it. Each model attempts the same tasks, so you can compare them.
3. Press Run agents →. You can close the page; we email you when the run finishes.

The agents call your server's real tools. A tool that creates, changes or deletes data will really do so. The agents name what they create mcpt-<run id>-… so you can find it, avoid tools that send messages, invite people, pay or publish, and clean up after themselves. Still, test against an account with test data. The report lists anything it knows was left behind.

You can run one test at a time on each server. Stop run ends a run at once. A stopped run has no report, and nothing is cleaned up after it.

Each agent answers one question about your server, using its own list of checks.

Agent	The question it answers	Checks
contract	Does each tool honour its own declared contract?	4
happy-path	Can an assistant do a real user's tasks with it?	4
robustness	Does the server fail cleanly instead of lying or hanging?	4
security	Truthful annotations, no model-aimed text, sound auth?	5
openai-store	Would it pass ChatGPT App Directory review?	5
anthropic-store	Would it pass Claude Connectors Directory review?	6

happy-path is the closest to real use. It writes tasks a user would bring to your server, then hands each one to an assistant model that sees nothing but your tools. Afterwards it checks whether the model really got the job done.

A list of findings, not a score. Each finding names the tool, what went wrong, the evidence, and a suggested fix.

Mark	Severity	Meaning
×	high	Breaks real use, or is a security problem. Fix first.
!	medium	Works, but misleads assistants or fails some of the time.
·	low	Polish: naming, descriptions, small gaps.

An agent that found nothing, or sent no report, is listed under Nothing to act on instead of getting a section of its own.

• happy-path's task list shows each task, worst first, with how many of the models you picked passed it. A task can pass, fail because of the server, fail because of how the tools are designed, or stay unverified.
• Left on your server appears only when the agents created test data they couldn't remove, and lists it so you can delete it.

Writes that ran but that no tool could read back to confirm are listed in the full report.

The full report has its own page, linked from the run's heading and from the server's list of runs. There you can save it as a PDF or copy it as a prompt for a coding agent.

Only the members of its workspace, until someone shares it.

• While a run is going, only the person who started it can open it. Once it finishes, every member of the workspace can.
• Sharing a finished run makes its report's link work for anyone, without signing in. On the report page, press share, then Copy link. Anyone with the link sees the report and nothing else: no link to the server's page, your other runs or your workspace.
• Stop sharing, in the same place, stops the link working at once.
• Any signed-in member can share a run or stop sharing it. A guest can't share a run.

A report quotes what your server returned. Read it before you share it.

A workspace holds servers, their runs, and the balance that pays for them. You can belong to several, and switch between them from the chip in the header.

• Owners rename the workspace, invite and remove people, and change roles.
• Members connect and delete servers, start runs, and see and delete every finished run.

To add someone, invite them by email from the workspace's dashboard. An invitation lasts 14 days, and it can only be accepted by a Google account with that email address. Anyone can leave a workspace. If the last owner leaves, the longest-standing member becomes the owner.

A base fee plus a fee for each agent you pick, shown on the run page before you start it. Runs are paid for from the workspace's balance, in US dollars.

• Happy path includes one standard model; each extra model adds a fee, and a flagship model (Claude Opus, GPT Sol) adds more.
• Your first run is free while today's free runs last, guests included: one agent with one standard model. Signing in keeps the free run you already used as a guest.
• An owner adds funds from the workspace's dashboard, or from the run page when the balance is short. Stripe takes the payment and sends the receipt.
• A run is charged when it starts. If a paid run fails before it shows any results because our service or our AI provider broke down, its price comes back to the balance. A run that uses up its budget, ends without its final report or reaches its time limit stays charged, whatever the cause. A free run that fails is not given again.
• The dashboard's ledger lists every amount added or spent, and invoices opens your receipts.

Yes. You can connect a server, read its tools and run a test without signing in. We keep your work in a guest session in this browser.

• A guest has one server in view at a time, and no dashboard, teammates or run history.
• When you sign in, you choose what happens to your guest work: move it into one of your workspaces, or leave it behind. Leaving it behind deletes it.
• Guest work you never sign in to keep is deleted 30 days after your last visit.

Each action erases what its line below says, and everything else stays with the workspace.

• Disconnect removes your login to a server. The server, its runs and your teammates' logins stay.
• Deleting a server erases it, every run on it and every login to it, for the whole workspace. Its links, public ones included, stop working. You can't delete a server while a run on it is going.
• Deleting a run erases its report and log for the whole workspace. Its link stops working. You can delete a run once it has finished, a teammate's included.
• Leaving a workspace removes your access and your logins to its servers. Its servers and runs stay with the workspace.
• Deleting your account erases your profile, picture, memberships and logins. Runs you started in shared workspaces stay with those workspaces and stop naming you.

Workspaces can't be deleted from the app yet. Write to privacy@mcptester.dev to have one erased.

500 servers and 750 members per workspace, 1,000 tools read from one server, and one run at a time per person per server. The full list is below; the ceilings are generous on purpose, and a real workspace shouldn't meet them.

What	Limit
Servers per workspace	500
Tools read from one server	1,000
Members per workspace	750
Workspaces you can create	100
Invitation emails	50 a day per workspace, 100 a day per person
Runs going at once	One per person per server
Connecting to a server	30 seconds
Contacting servers (connecting, server pages, Refresh)	Paced per network address. If asked, wait a minute

Email support@mcptester.dev for help with the product, a run or a server. For anything else:

• Questions about your data, or a request to erase something: privacy@mcptester.dev
• A security issue in MCP Tester: security@mcptester.dev
• Anything about the terms: legal@mcptester.dev

When you write about a run or a server, include the link to its page. How we handle your data is in the Privacy Policy, and the rules of use are in the Terms of Service.