Auth wired by hand, per client
Every MCP client wants its own configuration dance. Tokens get pasted into config files and end up in someone's dotfiles repo.
For agent and application builders
Ship the agent, stop hand-rolling the guardrails
Auth per client, redaction, tool scoping, and spend limits are the same 4 problems on every project. Put them in the proxy once and get on with the part that is actually your product.
The problem
What tends to keep you up
PII reaching the provider
A customer record ends up in a prompt because nothing sat between your code and the API to notice or stop it.
Tools that do more than intended
An MCP server exposes twenty tools. Your agent needed two. The other eighteen are reachable by anything holding a valid token.
A retry loop with a credit card
An agent gets stuck, retries at machine speed overnight, and the first signal anybody gets is the invoice at the end of the month.
What you get
The 4 problems, solved at the proxy
Clients that configure themselves
OAuth 2.1 with RFC 7591 Dynamic Client Registration and RFC 8414 discovery, with full PKCE. Cursor and Claude Desktop authenticate and start working without manual setup.
Read the guideTools off by default
Tools are discovered automatically but disabled until you enable them. A 3-tier hierarchy of user override, then group, then global decides who may call what, with sliding-window limits per user, proxy, and tool.
Read the guideRedaction that round-trips
PII and secrets are masked before the request leaves and the original values are restored in the response. Your users see real data, the provider never does, and your code does not change.
Read the guideA budget that actually stops
Monthly USD limits per team or per API key with a warning threshold and a hard-block option. Past the limit the proxy returns HTTP 402 instead of forwarding, so the loop stops costing money immediately.
Read the guideA playground before you commit
Exercise proxy authentication end to end, including the full OAuth flow, from inside the dashboard. Confirm a client will connect before you wire it into anything real.
Read the guideSystem prompts as configuration
Inject behavioural rules and security instructions into requests automatically, selected per proxy or per user group, with template variables. Change the rules without redeploying the agent.
Read the guide
Getting started
Your first half hour
Nothing here is irreversible, and every control is opt-in. You can run the whole sequence against a throwaway proxy before you point anything real at it.
-
01
Create an MCP proxy
Point it at the upstream MCP server you already use. The gateway discovers the available tools automatically and leaves every 1 of them disabled.
-
02
Connect your client
Point Cursor or Claude Desktop at the proxy URL. Dynamic Client Registration handles the rest, so there is no token to paste anywhere.
-
03
Enable only what the agent needs
Turn on the two tools it actually calls, and set a per-minute limit on the expensive one. Everything else stays unreachable even with a valid token.
-
04
Add a budget and redaction
Set a monthly cap on the key and turn on redaction for the fields you care about. Then leave the agent running overnight without worrying about it.
Stop rebuilding the same 4 things
Create a proxy, point a client at it, and see the tool list come up empty until you say otherwise.