Technical writing about MCP authentication, agent identity, and production auth.
June 30, 2026
X shipped an MCP server. OAuth says who the user is — nothing says what the agent can do. We wrapped xurl MCP with Bolyra Shield and ran three scenarios: attacker blocked, full agent authorized, delegated agent scoped to read-only.
Read post →June 22, 2026
The full credential lifecycle: Poseidon5 commitment, on-chain Merkle tree enrollment, off-chain resolution, tree-level revocation, circuit-enforced expiry, and key rotation via revoke-and-re-enroll.
Read post →June 22, 2026
OAuth identifies the user. Nothing identifies the agent. We put @bolyra/gateway in front of a mock Robinhood MCP server and ran four scenarios: verified read, verified trade, unauthorized agent blocked, replay attack blocked.
Read post →June 18, 2026
MCP auth belongs in front of your server, not inside it. @bolyra/gateway is a reverse proxy that verifies agent credentials, enforces per-tool permissions, prevents replay, and emits signed audit receipts. One command. Zero auth code in your server.
Read post →June 14, 2026
Local demos prove plumbing. Production needs attribution. Five questions your MCP demo can't answer: agent identity, human delegation, per-tool policy, replay protection, and audit receipts.
Read post →June 13, 2026
Dev mode skips 4 things that matter in production: credential resolution, Merkle root validation, nonce replay protection, and proof-to-credential binding. Here is the config diff and the fix for each one.
Read post →