The core principle stays the same:
An agent seeing a message is not authority to act on it.
ActionLock should make the transition from observation to execution explicit, reviewable, and evidence-bound.
ActionLock is live, but this is the first usable boundary, not the final form.
Next areas we are working toward:
Public key publication and rotation so receipt identities can be pinned outside a local installation.
Better retained-message observation, while staying honest about upstream history limits and unknown evidence.
More real-world adversarial tests around prompt injection, replay, policy drift, and downstream failure handling.
without exposing the local root secret.
Built for the Technocore ecosystem and agent builders who want a real boundary between โthe agent saw thisโ and โthe agent is allowed to do this.โ
https://t.co/Ug002sRI9A
GitHub: https://t.co/NEDdmWauup
#Technocore#MCP#AIAgents#FLOPLabs
AI agents can read messages from anywhere. That does not mean those messages should be allowed to choose tools, targets, or actions.
ActionLock is a local-first MCP capability gateway for agents consuming untrusted Technocore messages.
It separates:
evidence of what was observed
Approvals are short-lived, one-time, and consumed before execution.
Shell, wallet, and social actions requested by remote content remain blocked.
We also added separate Ed25519 receipts for the approved action and its eventual execution result, so the record can be verified