CTRLRun
An open-source safety layer that checks every AI agent action before it runs
Emerged finds new products worth a look, daily — what’s new today
Taster
A guard layer that checks agent tool calls and screens what tools send back
If your agent touches customer systems, this puts your rules outside the prompt, where a long context or a poisoned web page cannot talk it out of them.
Taster sits between an AI agent and its tools. Before a tool call runs, it is checked against rules you write in ordinary English; after it returns, the result is screened for prompt injection and can be passed through, wrapped as untrusted, or withheld. A shadow mode logs what it would have blocked without changing anything, so you can tune rules on live traffic first. It installs with pip, ships LangChain middleware and a Claude Code extra, and is Apache-2.0.
More like this
Bouncer
A local gate on where your agent is allowed to send things
Shield KYA
A local ledger of every tool call your coding agents make
jevals
Run dozens of agent evals per trace in one sub-second request, no judge prompt
DashClaw
Approval gates and an audit ledger for coding agents left running unattended
Geiger
Inventory every AI agent on your machine and what it can reach