yelno.aiChecking statusSign inGet API key

Every tool your agents need, behind one URL.

Connect the MCP servers your team uses once, with managed OAuth and per-user accounts. Group their tools into toolsets that Claude, Cursor, ChatGPT or your own agent reach at one URL, with approvals, rate limits and a record of every call.

  1. CLIENT
    Claude, Cursor, ChatGPT, your agentOne URL per toolset. OAuth or a Yelno key.
  2. TOOLSET
    Your rulesAllowed tools, read-only mode, approvals, rate limits.
  3. GATEWAY
    Every call governedCredentials attached, secrets redacted, call traced.
  4. SERVERS
    Only the ones you connectGitHub, Linear, your own: remote, over HTTPS.
What it does

Managed MCP, from connection to call.

Connect the MCP servers your agents need once. Decide exactly which tools they can use, who approves the risky ones, and see every call. The gateway runs tools only on servers you connect and allow.

01 · CONNECT

A directory of verified servers

Pick a server from a directory that platform staff review, or add your own by URL. Connect once for the whole organization, with OAuth or an API key, and every tool it exposes is listed with its annotations.

02 · AUTHORIZE

Managed OAuth

Yelno runs the server's OAuth flow with PKCE, stores the tokens encrypted and refreshes them before they expire. Your agents and clients never hold the server's credentials, and a connection that needs re-authorizing says so.

03 · PER USER

Connected accounts for your customers

Connect a server on behalf of each of your own end users, keyed by your external user id. Send them a connect link, and the gateway calls tools with that user's account, never another's.

04 · COMPOSE

Toolsets with one URL

Choose connections and exactly which tools to expose, turn on read-only mode, require approval for anything destructive, and set rate limits. Each toolset is its own MCP server, with one URL for any client, signed in with OAuth or a Yelno API key.

05 · CONTROL

Approvals and safety

Calls that need a person wait in an approvals inbox: approve once, deny, or always allow that tool. Private addresses are refused, secrets in tool output are redacted, and likely prompt injections are flagged.

06 · OBSERVE

Every call on the record

Calls over time, error rate and latency per connection and per tool, health checks for every server, and a recent-calls log you can filter. Tool arguments and results are kept only when your organization's content capture policy allows it.

Directory

Verified servers, ready to connect.

Platform staff check each entry against the vendor's own documentation before it is listed. Anything not in the directory can still be added by URL.

The full directory

The directory covers developer tools, project management, knowledge bases, data, search and cloud platforms, each with its sign-in method. Browse every verified server, or connect your own by its URL from the console.

Works with

One URL for any MCP client.

Each toolset is an MCP server over Streamable HTTP. Clients that support OAuth sign in on a Yelno consent screen; everything else uses a Yelno API key.

Claude Desktop and claude.ai
Add the toolset URL as a custom connector, signed in with OAuth or an API key header.
Cursor
Add it to .cursor/mcp.json as a remote server with an Authorization header.
VS Code
Add it to .vscode/mcp.json as an http server; the editor prompts for the key once.
ChatGPT
Add it as a connector in developer mode, with OAuth.
OpenAI Responses API
Pass {"type": "mcp", "toolset": …} in tools; the gateway runs the tool calls and bills the model as usual.
Anthropic SDK and any MCP client
Point the MCP connector or any Streamable HTTP client at the toolset URL with your Yelno key.
.cursor/mcp.json
{
  "mcpServers": {
    "yelno-eng": {
      "url": "https://yelno.ai/v1/mcp/toolsets/mcpts_…"
    }
  }
}
Responses API
client.responses.create(
    model="MODEL",
    input="Summarize the open issues assigned to me.",
    tools=[{"type": "mcp", "server_label": "eng",
            "toolset": "mcpts_…"}],
)
Approvals & safety

Risky calls wait for a person.

Destructive tools need approval by default. A client signed in with OAuth on the current MCP version can ask its user in the chat; any other client gets a link to the approvals inbox, and the call runs once it is approved.

Only servers you connect

The gateway calls tools only on remote MCP servers your organization has connected and enabled. It does not run code, local servers or tools of its own choosing.

New hosts are a policy decision

Registering a server lets the gateway send data to a new host, so it follows your organization's approval policy for relaxing data rules. Platform staff can also block a host for everyone.

No private networks

Private, loopback, link-local and cloud-metadata addresses are refused when the connection is made, redirects are never followed, and responses are size-capped.

Destructive tools wait

By default a tool needs approval unless its server marks it read-only or not destructive. Read-only mode exposes only tools marked read-only.

Secrets redacted, injections flagged

API keys, tokens and private keys found in tool output are replaced with a marker before the client sees it. Text that looks like a prompt injection is flagged on the call, never silently changed.

Credentials stay here

OAuth tokens and API keys for connected servers are stored encrypted and sent only to that server. They are never shown again, logged or sent to a model.

You remain responsible for the tools

The gateway cannot judge what a tool does with the data it receives. Connect only servers you trust, and review their tools before you expose them.

Pricing

Tool calls are free by default.

You pay for the models that call tools, at their usual final price. If a per-call price ever applies to your organization, the console shows it first, and a call is charged only when the server returns a successful result.

NO SUBSCRIPTION

No seat fee and no platform fee for the gateway. Every charge, for a model or a priced tool call, is held before it is spent and listed on your statement. See pricing for the models.

Step by step

From server to client in minutes.

Every step is in the console and is also an API route.

Full guide
  1. 01

    Connect a server

    Pick it from the directory, or add a custom server by URL. Registering a new host follows your organization's approval policy.

  2. 02

    Authorize it

    Sign in with OAuth or paste an API key for the whole organization, or send connect links so each of your users connects their own account.

  3. 03

    Build a toolset

    Choose the tools, set read-only mode, approval rules and rate limits, and scope it to projects or keys.

  4. 04

    Connect a client

    Paste the toolset's MCP URL into Claude, Cursor, VS Code or ChatGPT, or pass the toolset to the Responses API.

  5. 05

    Approve and observe

    Approve the calls that need a person, and watch calls, errors and latency per tool in the console.

QUESTIONS

What teams ask first.

The guide covers connections, toolsets and every client; the API reference lists every route.

What is the MCP gateway?
A managed layer between your AI clients and the MCP servers you use. You connect servers once, choose which tools are exposed in a toolset, and give any MCP client one URL. The gateway handles OAuth, approvals, rate limits and logging for every tool call.
Which servers can I connect?
Remote MCP servers over HTTPS: servers from the directory, which platform staff verify against each vendor's own documentation, and custom servers you add by URL. Local (stdio) servers are not supported. Registering a server follows your organization's approval policy.
Does the gateway run tools on its own?
No. It calls a tool only when a client or a model asks for it through one of your toolsets, only on a server your organization connected, and only if that tool is enabled. Calls that need approval wait until a person approves them.
How does OAuth work?
Both ways. When you connect a server that uses OAuth, Yelno sends you to the server's own sign-in page with PKCE and stores the tokens it returns, encrypted, refreshing them before they expire; if access is revoked the connection asks to be re-authorized. And clients such as Claude or ChatGPT can connect to a toolset with OAuth: Yelno shows a consent screen and issues a token valid for that one toolset.
Can my customers connect their own accounts?
Yes. Create a connect link for an end user, identified by your own external user id. They authorize the server themselves, and calls made for that user use only their account.
Which clients work with it?
Any client that speaks MCP over Streamable HTTP, current or earlier protocol versions: Claude Desktop and claude.ai connectors, Cursor, VS Code, ChatGPT connectors and the MCP connector in the Anthropic SDK. The Responses API also accepts a toolset or a connection directly.
What is recorded about a tool call?
Always the toolset, connection, tool, outcome, error code, duration and sizes. Tool arguments and results are kept only if your organization captures content for MCP, and never when it requires zero data retention. Credentials are never recorded.
What does it cost?
Gateway tool calls are free by default. If a per-call price applies to your organization, the console shows it, and a call is charged only when the server returns a successful result. Models that call tools are billed at their usual final price. There is no subscription.
Is there an uptime guarantee?
No. The gateway depends on the MCP servers you connect, which Yelno does not operate, so no availability or SLA is offered. Each connection's health is shown in the console.

Yelno does not operate the MCP servers you connect, so no availability or SLA is offered for them or for the gateway. Tool annotations such as read-only and destructive are the server's own hints. The gateway runs tools only on servers your organization connects and allows.