MCP · 5 min
What is an MCP server, explained for API teamsRead →MCP Management
MCP servers your security team will approve
Generate secure, governed MCP servers from your API contracts, put every tool call behind an authorization gateway, and see exactly which agent did what. Hosted, scoped, logged.
SOC 2 Type IIOAuth2 + per-tool scopesFull call logsSSO / SAMLOn-prem available
Claude
CursorPartner bot
Authorization gatewaymcp.elva.dev/acme-paymentsActive
auth OAuth2scopes per-toolrate 600/minlogs on
✓get_payment_statusClaudescope ok · 98ms
✓create_invoiceCursorscope ok · 145ms
✕delete_accountPartner botdenied · out of scope
Your API · payments-api v2.4.123 tools exposed of 214 endpoints
99.9%UPTIME SLA
138P95 GATEWAY LATENCY (MS)
EveryCALL LOGGED, NO SAMPLING
0UNSCOPED TOOLS ALLOWED
01MCP generation
Governed MCP servers, built from contracts and not hand-written glue
A contract already says which endpoints an audience can see and which fields they get. Elva turns that into a typed MCP server: curated tools, rich descriptions, output schemas, and an install page you can hand to a partner.
Tools exposedcurated from 214 endpoints23 of 214
✓get_payment_statustyped outputread:payments
✓create_invoiceidempotentwrite:billing
✓list_transactionspaginatedread:payments
!refund_chargeheld back: missing output schema
-191 endpoints not exposedout of contract scope
https://mcp.elva.dev/acme-paymentsInstall page
Curated, not dumpedAgents choose tools by reading descriptions. A 214-tool server makes every choice worse. Elva exposes only what the contract allows and what scores well enough to be understood.
Schema changes are gatedA tool schema is a promise agents take literally. Remove a field an agent depends on and the publish is blocked until someone approves it.
Hosted or your own cloudRun on Elva with a custom domain, or deploy inside your VPC on Enterprise. Same contracts, same logs, your perimeter.
02Authorization gateway
No agent gets a raw key to your API
Every tool call passes through the gateway: identity checked, scope enforced, rate limited, logged. Keys are issued per agent and revoked in one click, and your upstream credentials never leave Elva.
Keys and scopesIssue key
elva_live_...8f2aClaude Desktop
read:paymentsread:invoices600 req/min
elva_live_...c91dCursor IDE
read:payments120 req/min
elva_test_...44b0Revoked 2m ago
Elevated error rate, 5.8% · scope escalation attempt blockedPolicy checks per callin order, fail closed
01IdentityOAuth2 token or managed key, mapped to an agent
02Tool scopeIs this key allowed to call this specific tool
03Field redactionPII stripped from responses per contract
04Rate and quotaPer key and per tool, retry storms contained
05Audit recordAgent, key, tool, arguments hash, latency, outcome
03MCP Playground
Test your MCP server in the Playground, not in production
Chat with an agent against your own server and watch every tool call land in a live log: the arguments it sent, what came back, and how long it took. Playground calls do not count against your quota.
MCP Playgroundacme-payments v2 · 23 tools
Refund the duplicate charge on order 4821 and confirm it went through.
get_payment_status{ "orderId": "4821" }ok · 98ms
list_transactions{ "orderId": "4821", "limit": 20 }ok · 172ms
refund_charge{ "txnId": "txn_9f21" }denied · missing an output schema
Order 4821 was captured twice for $240.00. The duplicate is txn_9f21. I cannot refund it from this server: refund_charge is held back until its output schema verifies, so the gateway denied the call.
Ask the agent to use your MCP tools…
Claude Sonnet 4.6
LogsLive
get_payment_status98ms · 820 tokok
list_transactions172ms · 1,140 tokok
refund_charge4ms · 0 tokdenied · missing an output schema
A tool list is not a testA server can expose exactly the right tools and still fail, because the agent cannot tell which one to pick. The only test that settles it is whether a model reaches the answer.
The gateway is in the loopPlayground calls go through the same gateway your agents will. A tool the contract holds back is denied here too, so you find the gap before a partner does.
Run it as often as you likePlayground calls do not count against your tool-call quota, on any plan. Re-run the same prompt while you tighten a description and watch the picks change.
04MCP logs
Answer “which agent did this” in one query
Every call, not a sample. Filter by agent, key, tool, or outcome, export to your SIEM, and set retention to match your policy. When an auditor asks, the answer takes a minute.
Live call logstreaming
Export
TIMEAGENTTOOLKEYLATENCYTOKENSOUTCOME
14:32:12Claude Desktopget_payment_status...8f2a98ms820ok
14:32:10Cursor IDEcreate_invoice...c91d145ms1,340ok
14:32:08Partner botdelete_account...44b04ms-denied · out of scope
14:32:06Claude Desktoplist_transactions...8f2a172ms4,100ok
14:32:03Windsurfrefund_charge...7e15210ms960error · schema mismatch
14:32:01Partner botsync_ledger...44b02410ms2,200ok · slow
14:31:59Cursor IDEget_customer...c91d88ms640ok
retention 30 days on Business · custom on Enterpriseexport: JSON, CSV, webhook to SIEM
05MCP insights
See cost, latency, and blast radius per agent
One agent doing most of your traffic is either your best integration or a retry loop. Insights separate the two, surface the tools that are slow or failing, and show what tokens are actually costing you.
Tool calls (24h)847↑ 6.7% from yesterday
p95 latency138msstreaming included
Denied calls129 scope, 3 rate limit
Token spend (24h)124kavg 1,240 per call
Traffic by agent
Claude Desktop412 calls · 98ms
Cursor IDE233 calls · 145ms
Partner bot147 calls · 210ms
Windsurf55 calls · 132ms
Needs attention3 tools
!refund_charge14% errors · missing output schema
↑list_transactions4,100 tokens avg · unpaginated
◷sync_ledgerp95 2.4s · agents timing out
06Agent feedback
Agent feedback tells you why a call failed, not just that it did
A log says a call failed. Agent feedback says why: an unclear description, too many parameters, a server that timed out. Claude, ChatGPT, Cursor, and Windsurf file each one against the endpoint it came from.
Agent feedbackSummarize with AI
Filter by date5 open · 1 critical
ClaudeCritical2h ago
The tool description does not clearly explain when it should be used.GET/customers/{customerId}
ChatGPT4h ago
The tool failed to return a response after multiple attempts. I could not complete the request because the MCP server kept timing out.POST/orders
Cursor4h ago
The tool requires too many parameters for a simple request.GET/products/{productId}
Claude7h ago
Two tools have almost the same description, so I picked the wrong one twice before it worked.GET/orders/{orderId}
Windsurf9h ago
The response has no pagination cursor, so reading the whole list costs 4,100 tokens every call.GET/transactions
filed by the agents calling acme-payments v2scored into the catalog
Filed against the tool, not a mailboxEach item carries the client that hit it, the endpoint it came from, and when. Nobody has to work out which of 214 endpoints the complaint was about.
Wording is a bug reportAn agent that cannot tell two tools apart is describing a defect in your descriptions, not a bad day for the model. Elva treats it as one and ranks it by how often it blocks a call.
It ends up in the catalogFeedback lands on the endpoint next to its AI readiness score and the suggested fix, so the next contract version ships with the description rewritten and the parameters trimmed.
Enterprise readiness
Secure MCP servers, built for the review your platform team runs
Servers run on SOC 2 Type II and ISO 27001 certified infrastructure. Every control a security team asks about before approving an MCP server is covered in full on the MCP server security page.
Least privilege by defaultTools ship scoped or they do not ship. No wildcard keys, no ambient access to endpoints outside the contract.
Your credentials stay putUpstream API secrets live in the gateway, never in an agent config or a partner’s machine. Rotate without redistributing anything.
Revoke in secondsOne click kills a key, an agent, or a whole server. The log keeps everything it did up to that moment.
SSO, SAML, SCIMYour identity provider owns who can publish and approve. Deprovisioning a person removes their access everywhere.
Audit trail and approvalsWho changed which contract, who approved the breaking change, and when the server picked it up. Exportable.
Private cloud or on-premRun the gateway inside your network with a custom domain, keep logs in your account, and satisfy data residency.
FAQ
What is the Elva MCP gateway?
The hosted layer agents connect to. Every tool call passes five checks in order: identity, tool scope, field redaction per contract, rate and quota, and an audit record with actor, tool, arguments hash, latency, and outcome.
Can Elva generate an MCP server from my API?
Yes. MCP servers are generated from a contract view of your catalog, so agents only see the tools an audience was deliberately granted. Servers are hosted by Elva, or self-hosted on Enterprise.
Which AI clients can connect?
Any MCP client, including Claude, Cursor, and ChatGPT, plus partner agents using scoped keys through the same gateway.
How many tool calls are included?
1,000 per month on Free, 25,000 on Startup, 250,000 on Business, and custom on Enterprise. Paid plans continue past the limit with metered overage at $10 per 10,000 calls.
How do I test an MCP server before agents use it?
In the MCP Playground. Point an agent at your own server, watch every tool call it makes with the arguments, the response, and the latency, and see which calls the gateway allows. Playground calls do not count against your tool-call quota.
How do agents report problems with my API?
Through agent feedback. When a tool is hard to use, the agent files a report against the endpoint it came from: an unclear description, too many parameters, or a response that timed out. Elva scores it and turns it into a fix in your catalog.
- SOC 2 Type II · certified
- ISO 27001 / 27701 · certified
- GDPR · compliant
- Encryption · TLS 1.3 / AES-256
- SSO / SAML and SCIM · available
Give agents a door, not a spare key
Deploy a scoped, logged MCP server from your first contract in minutes.