A trust score an AI agent can check before it connects
The AdminKit Agent Trust API lets an AI agent ask about an unfamiliar MCP server before connecting: one GET request returns a 0-100 score, a verdict and the evidence behind it, such as whether the server answers without authentication or exposes high-risk tools anonymously, so framework maintainers get a documented reason a call proceeded or stopped.
Updated
At a glance
initialize, anonymous tool listing, high-risk tool names and CORS policy.GET /v1/trust/mcp?url= and GET /v1/trust/endpoint?host= on api.adminkit.net, listed in the OpenAPI reference.What problem does the Agent Trust API solve?
AI agents are increasingly built to discover and call tools they haven’t seen before, sometimes at runtime, based on what a task requires. That’s powerful, but it also means an agent can end up connecting to an endpoint with no prior knowledge of whether it’s trustworthy.
A human developer reviewing a new integration usually does some due diligence first: checking certificates, reading documentation, looking at who operates it. An autonomous agent making that decision in milliseconds has no equivalent step unless one is built in.
Without some kind of check, an agent either has to blindly trust every endpoint it’s pointed at, or a developer has to hardcode an allowlist in advance, which defeats the point of letting the agent discover tools dynamically.
Who is the Agent Trust API for?
Agent framework maintainers
Add a built-in check before tool calls, so every agent built on your framework gets a reason to proceed or stop.
See the solution →
Developers of autonomous agents
Let an agent discover MCP servers at runtime without hardcoding an allowlist in advance.
See the solution →
Security reviewers
Keep a documented reason for why an agent connected to a server, not a black-box decision.
See the solution →
How does the Agent Trust API work?
Before making the real call, an agent (or the code orchestrating it) asks the Agent Trust API about the MCP server it’s about to use. AdminKit probes that one endpoint the way an anonymous client would, following the Model Context Protocol specification, and returns a score from 0 to 100, a verdict, and the findings behind it. The agent or its developer then decides whether to proceed, proceed with limits, or stop.
The findings map to risks named in the OWASP Top 10 for LLM Applications, such as excessive agency through tools anyone can call.
An agent meets a new server
Your agent or orchestrator is about to connect to an MCP server it hasn't used before.
Ask for a score
GET /v1/trust/mcp?url= with the server URL, or /v1/trust/endpoint?host= for a bare host on the /mcp path.
AdminKit probes it
AdminKit connects as an anonymous client would and checks authentication, anonymous tool listing, high-risk tool names and CORS.
Your policy decides
Proceed, limit or stop on the 0-100 score and verdict, and keep the findings as the documented reason.
curl -s 'https://api.adminkit.net/v1/trust/mcp?url=https://mcp.example.com/mcp'
# Every response has the same JSON shape (TrustScoreResponse in the OpenAPI reference):
subject https://mcp.example.com/mcp
score 0-100
status trusted | caution | untrusted | unknown
evidence summary, findings (id, severity, title)
checked_at when the probe ran
What do you get with the Agent Trust API?
A 0-100 score
One number an agent's policy can compare against a threshold before it connects.
score
Four clear verdicts
Trusted, caution, untrusted, or unknown when the probe could not reach a verdict, for example because the server was unreachable.
status
The evidence behind the score
Findings such as no authentication required, tools listed anonymously, a high-risk tool exposed, or an unsafe CORS policy.
evidence.findings[] (id, severity, title)
Ask by URL or by host
Score a full MCP URL, or a bare host on the conventional /mcp path.
GET /v1/trust/mcp?url= and /v1/trust/endpoint?host=
No key to start
Anonymous calls are accepted and rate-limited per IP, the same way as the free tools. An API key raises the limits.
X-API-Key header, RateLimit-* response headers
Public endpoints only
Private and internal addresses are refused, so the API can't be used to probe someone's internal network.
Private-network targets rejected
What does a response contain?
| Field | Meaning |
|---|---|
subject | The URL or host you asked about |
score | 0-100, from the probe’s findings |
status | trusted, caution, untrusted or unknown |
evidence | The findings (id, severity, title) and a summary of what the probe saw |
checked_at | When the probe ran |
A score reflects what AdminKit observed when the endpoint was checked, not a continuous guarantee. Private and internal addresses are refused, so the API only probes public endpoints. Need trust scoring wired into your own agent platform, or signals beyond MCP? Talk to our team.
Which findings lower the score?
| Finding id | What the probe saw |
|---|---|
mcp.no_auth_required | The server answered initialize without any authentication |
mcp.tools_exposed_anonymously | The tool list was readable by an anonymous client |
mcp.dangerous_tool_exposed | A listed tool’s name suggests shell, file or network access |
mcp.unsafe_cors | CORS allows any origin together with credentials |
The same probe runs on the free mcp-check page, so you can see a server’s findings in a browser before wiring the API into an agent.
What does the Agent Trust API work with?
Standards it speaks
- Model Context Protocol
- HTTPS and JSON
- OpenAPI 3.1 contract
Using and able to integrate
- Any agent framework that can make an HTTPS call
- The free mcp-check tool page (same probe)
Who builds and supports the Agent Trust API?
AdminKit is a product of Netbase JSC, a software company founded in 2012 that delivers its projects remotely. The same team that wrote the Agent Trust API runs the API and answers your questions. Read more about Netbase JSC and AdminKit.
What Netbase JSC's clients say
Reviews of Netbase JSC's project work, not of AdminKit — AdminKit is a newer product from the same company.
"They had nice and good solutions for all of our requests."
Daniel Popescu
Owner, Digital Art Studio · Romania
Source: Clutch"They’re good listeners, are understanding, and they remember things very well."
Syed Ali
Partner, Oriental Overseas Exports · India
Source: Clutch"Their Quick response in every step was impressive."
Olof Christensson
CEO, Skelderwikens Brygghus AB · Sweden
Source: ClutchCan Netbase set the Agent Trust API up for you?
The Agent Trust API is self-serve, but you don't have to wire it in alone: Netbase can connect it to your pipeline, agent platform or internal tools as a scoped engagement.
Netbase expert service
MCP / agent hardening
A Netbase review of a public or internal MCP server: authentication, tool allowlists and prompt-injection-aware checks, or trust scoring wired into your own agent platform.
Does the Agent Trust API fit a larger organisation?
- Trust scoring wired into your own agent platform, built by Netbase.
- Signals beyond MCP, scoped and built as custom development.
- Custom call volumes, quoted by our team.
Talk to an expert
Huy Nguyen (David), Founder & CEO, Netbase JSC
Ask about the Agent Trust API for your organisation: scope, timeline or a quote.
What do people ask about the Agent Trust API?
How do I call the Agent Trust API?
Send GET https://api.adminkit.net/v1/trust/mcp?url=<the MCP server URL>, or GET /v1/trust/endpoint?host=<host> to check a bare host on the conventional /mcp path. No API key is needed for small volumes; calls are rate-limited per IP like the free tools.
What does a trust score mean?
It's one input into a decision, built from what AdminKit observes when it probes the endpoint: whether it requires authentication, whether its tool list is readable anonymously, whether a high-risk tool name is exposed, and how its CORS policy is set. It is not a guarantee that an endpoint is safe.
What verdicts can the API return?
trusted, caution, untrusted or unknown. Unknown means AdminKit could not reach a verdict, for example because the endpoint was unreachable — that is different from finding something wrong.
Does a low score mean an endpoint is malicious?
Not necessarily. A low score usually means the endpoint is missing common safeguards, such as authentication on tool calls. It is not proof of malicious intent.
Does AdminKit vouch for endpoints with a high score?
No. AdminKit reports what it observed at the time of the check; it does not vouch for or guarantee the behavior of any third-party endpoint.
Where should you go next?
All AdminKit products →Product
MCP & Agent Gateway for AI Tool Calls
Netbase builds and runs an MCP gateway for you: OAuth without token passthrough, tool allowlists, rate limits and an audit log of every call.
Solution
Check what your AI agents and MCP servers expose, before you ship
Check if your MCP server is reachable and needs auth for free with AdminKit. Score endpoints with the Agent Trust API, or have Netbase run a gateway.
Guide
How do you read an MCP endpoint check result?
What an MCP endpoint check's initialize and tools/list results mean, including the SSRF guard, and what to do about an unauthenticated server.
Start using the Agent Trust API
Small volumes work without a key. Create a free account for an API key, or ask our team about custom volumes and a dedicated deployment.
How AdminKit works with you
- Delivered remotely — Netbase JSC, the company behind AdminKit, has worked with clients in 80 countries.
- Contracts and invoices are issued by Netbase JSC; payment goes to the company by bank transfer to Netbase JSC's company account or PayPal Business.
- Talk to the CEO, Huy Nguyen (David), directly on WhatsApp or by email at [email protected].