kubbi

kubbi targets a narrow but painful problem: passing payloads between AI agents without leaking sensitive data into logs, queues, or webhook histories. For teams building multi-agent architectures with MCP, A2A, or custom microservices, it provides a clean claim-check pattern out of the box. When to pick this: you need secure, ephemeral payload transfer across trust boundaries (e.g., different vendors, different teams). A common example is an MCP server returning a claim URL instead of embedding result data inline — the consuming agent can fetch the payload only if authorized. The free tier supports up to 20 kubbis per day with 16 KB payloads, enough for prototype evaluation. When to pass: your payloads exceed 25 MB (even on Enterprise), you need real-time streaming, or you require consumer-side authentication (kubbi is URL-based only). Also avoid if you need long-term storage — payloads auto-expire based on TTL. Compared to alternatives like S3 presigned URLs, kubbi adds server-enforced burn-after-read and multi-file bundling without requiring IAM setup. Compared to inline queue messages, kubbi avoids 256 KB SQS limits and keeps sensitive data out of logs. For teams that would otherwise build their own claim-check service with KMS and Lambda, kubbi saves that weekend of plumbing. Real-world caveats: the free tier's 2-hour max TTL and 5 active kubbis at a time can be limiting for production. The consumer side has no authentication, so the claim URL must be kept secret. Payload size limits (25 MB on Enterprise) may not fit large media files. Rate limits of 20 req/min on free tier can pinch during testing.