Pinned Tweet
We're running a survey on MCP adoption and would love your input. Takes ~5 minutes, and we'll share the full results with all respondents.
zuplo.link/mcpsurvey
Whether you're building servers, integrating MCP, or just exploring it, your perspective matters.
@MCP_Community
Prompt injection usually arrives in what a tool returns, not the user request. That is an outbound check. Zuplo's prompt-injection-outbound policy screens responses before an agent reads them, and strict defaults to false, so detection failures pass: zuplo.com/blog/protect-mcp-a…
The 2026-07-28 MCP spec deprecates OAuth Dynamic Client Registration in favor of Client ID Metadata Documents, and binds client credentials to the issuer that issued them. Per-server auth means per-server rewrites. Keep identity in your IdP: zuplo.com/blog/mcp-gateways-…
GitHub's secret scanning partner program can alert you when your API keys leak, but only if they carry a unique prefix and a checksum. That format decision cannot be retrofitted without rotating every customer key. Nine practices, walked through: zuplo.com/videos/api-key-bes…
ALT Don't build an API Key service until you watch this!
A User-Agent header is a self-reported label. Anything can claim to be a browser.
Sept 1: the IETF Web Bot Auth WG adopted HTTP Message Signatures for automated traffic. Identity from a signature you verify, not a header you trust: zuplo.com/videos/identify-ai…
ALT Identify AI Agents Using HTTP Message Signatures
Switching AI providers shouldn't mean rewriting your apps. The Zuplo AI Gateway now supports Bedrock Mantle, Grok, Gemini, and Mistral — swap models without touching app code, run multiple configs of one provider, keep spend attributed per app. zuplo.com/blog/attribute-ai-…
"We decreased by probably a factor of ten the amount of time we had to invest in working with our API gateway." — Tyler Hall, Software Engineer, Spot AI. They also cut API management costs 80% after leaving Apigee. zuplo.com/customers/spotai
Enterprise MCP auth shouldn’t mean rebuilding OAuth inside every connector.
With Zuplo’s MCP Gateway, authMode: "id-jag" exchanges an IdP identity assertion for a scoped upstream token—without a second user-consent loop. zuplo.com/mcp-gateway
New in the AI Gateway: DLP with 80+ detection rules (PII, credentials, financial data), Akamai AI Firewall support, and locked baseline policies every new app inherits. zuplo.com/blog/attribute-ai-…
Claude’s enterprise-managed auth for MCP connectors is generally available: centralized through your IdP, without individual OAuth per user.
Zuplo’s MCP Gateway already supports the ID-JAG flow behind enterprise connectors. xitter.cf/claudedevs/status/2091…
Models get retired. Agent adoption shifts. Infrastructure should survive both.
An agnostic MCP gateway lets teams swap Claude, Codex, Cursor, Copilot, or whatever comes next without rebuilding auth and governance.
zuplo.com/blog/bring-your-ow…
Your coding agent will change. Your MCP governance shouldn’t.
Keep auth, observability, and guardrails portable across Claude, Codex, Cursor, Copilot, and whatever comes next.
Why “bring your own agent” matters:
zuplo.com/blog/bring-your-ow…
Your AI invoice shows what you spent — not who spent it. New in the Zuplo AI Gateway: attribute usage to apps and teams, set spend limits at every level, fall back to cheaper models at the cap. zuplo.com/blog/attribute-ai-…
Zuplo retweeted
A protocol for agents to self-register dropped last week! 🤯
In one week: @resend , @NanoGPTcom , @Zuplo , @webclaw_io , @Cloudflare all shipped it. Most people I talk to haven't heard of it yet.
Zuplo retweeted
Zuplo now supports Cross App Access (XAA), the enterprise authorization extension for MCP.
What this means for your builds:
✅️ OAuth-based, open protocol with no vendor lock-in
✅️ Plug into Okta's central policy engine directly
✅️ Short-lived, scoped tokens replace hardcoded credentials
✅️ Full audit trail out of the box
XAA is now natively supported in TypeScript and Java MCP SDKs. Python support is coming soon.
zuplo.link/xaa
#MCP #OAuth #DevSecurity #EnterpriseAI
Your engineer's agent and your support rep's agent see the same MCP tools, including delete files and merge PRs.
Two gateway routes, one upstream, gated by IdP group:
zuplo.link/PszwKEb
Three LLM providers = three SDKs, three credential sets, zero fallback when one 500s.
14 things devs actually want from an AI gateway, ranked from Reddit threads:
zuplo.link/DbzWFRW
ALT What Developers Want from an AI Gateway
Swapping from GPT-4o to Claude shouldn't touch eight files, two SDKs, and a credential rotation. 14 things devs want from an AI gateway, ranked honestly:
zuplo.link/JJNekMZ
ALT What Developers Want from an AI Gateway
Your team set up an MCP registry and called it governance. But a registry is a catalog, not a control plane. It can't scope permissions by role or log what agents actually called. Know the difference:
zuplo.link/tcmihIE
40% of MCP servers need no auth. 43% have command injection vulns. 79% pass credentials through plain env vars.
Seven flaws to audit before your agent stack becomes an incident:
zuplo.link/zmap7ag
ALT 7 Common MCP Security Flaws and Vulnerabilities
You connect an agent to an MCP server to read data. The same tool list also lets it write, delete, and refund.
Nothing sits between the read and the destructive call but the model's judgment.
Front it with a gateway, allowlist only the tools you trust:
zuplo.link/5oeupl1
ALT Expose Only the MCP Tools You Choose