Superagent
open-sourceSummary
Superagent is an open-source SDK, released under MIT, that adds safety checks to AI agents at runtime, blocking prompt injections and unsafe tool calls, redacting PII, PHI, and secrets from text, and scanning repositories for agent-targeted attacks like repo poisoning. It ships as TypeScript and Python client libraries that call out to a guard/redact/scan service, so teams wire it directly into an agent's request path rather than deploying it as a standalone appliance. It is built for developers building or hardening AI agents rather than for traditional SOC or network security staff, sitting in the emerging ai-security category alongside runtime protection and data-leak-prevention tools. First committed in 2023 and backed by Y Combinator, it has drawn 86 contributors and 241 commits in the past year, indicating active development.
What Superagent answers
What does a blocked request look like to my code?
the guard call returns a classification plus the specific violation types that triggered it, so the calling code decides what to do next
Can I try this without committing to wiring it into an agent's request path?
the same client can scan a public repository URL directly and return a security report, independent of any running agent
Does it tell me what a scan costs?
each scan result includes a usage cost figure alongside the report, so repeated scanning against many repos is something you can meter
Do I have to choose between blocking bad input and cleaning sensitive output?
guard and redact are separate calls on the same client, so you can use either independently depending on whether the risk is an incoming instruction or outgoing text
Does this only work with one language's agent stack?
client libraries exist for both TypeScript and Python with matching methods, so a mixed-language agent codebase isn't split across two different tools
Examples
Command line
No option matches that search.
| option | found in | since | description |
|---|
No option matches that search.
Values are placeholders taken from each option’s declared default. Nothing is executed here — the output shown is a recording of a run that already happened.
Release history
Nothing shipped since we started watching. Releases and docs updates land here as they happen.