Connecting to the MCP server
Endpoints
| Host | Purpose |
|---|---|
https://mcp.theneuralledger.com/mcp | Canonical endpoint. Use this. |
https://mcp.tnl.st/mcp | Short transport alias. Same server. |
Both hosts use the OAuth resource identifier https://mcp.theneuralledger.com/mcp as the token audience, regardless of which host you connect through.
Authentication
The MCP server uses OAuth 2.0, not API keys. A token must satisfy all of:
- valid signature
- expected issuer
- audience
https://mcp.theneuralledger.com/mcp - unexpired
- scope includes
tnl:read
Requests failing any of these are rejected before any tool runs.
Client registration
Named OAuth clients require a static allowlist entry on the server side — contact support to have your client registered.
Dynamic client registration (DCR) is supported only when the issuer's DCR policy restricts callbacks and scopes, and only when the server has TNL_MCP_AUTH_ALLOW_DYNAMIC_CLIENTS=true. In that mode the gateway accepts issuer-generated UUID client ids. Signature, issuer, audience, expiry, and tnl:read validation still apply in every case.
Client configuration
Most MCP clients accept a remote server URL and handle the OAuth handshake interactively.
{
"mcpServers": {
"tnl-intelligence": {
"type": "url",
"url": "https://mcp.theneuralledger.com/mcp"
}
}
}claude mcp add --transport http tnl-intelligence https://mcp.theneuralledger.com/mcp{
"name": "tnl-intelligence",
"transport": "http",
"url": "https://mcp.theneuralledger.com/mcp",
"oauth": {
"scope": "tnl:read",
"resource": "https://mcp.theneuralledger.com/mcp"
}
}Your client will open a browser for the OAuth consent step on first connection, then cache the token.
Timeouts
Research tools do real work and can take a while. The budget is:
| Layer | Limit |
|---|---|
| Application | 180 seconds |
| nginx proxy | 190 seconds |
| Node request | 190 seconds |
Configure your client's timeout to at least 180 seconds. Clients defaulting to 30 or 60 seconds will abandon legitimate long-running research.
Usage accounting
MCP tool calls consume your monthly allowance, and every response carries a usage block reporting exactly where you stand:
{
"usage": {
"month": "2026-07",
"calls": 412,
"remaining": 9588,
"monthlyLimit": 10000,
"resetAt": "2026-08-01T00:00:00.000Z"
}
}A tool call that exceeds the allowance fails with a quota error rather than returning partial research.
Verifying the connection
Once connected, the client should list six tools prefixed tnl_research_. If it lists none, the token is missing the tnl:read scope. If the connection drops around the one-minute mark, the client timeout is too low.