Let your agents read the internet without taking orders from it.
Giving an AI agent web access means every page it reads could try to instruct it. ReachKit, a hardened rebuild of an MIT-licensed open-source toolkit, makes the defence structural: content from the internet always comes back labelled as untrusted.
- Type
- Python library
- Status
- Working scaffold: security core done and tested
- Stack
- Python 3, pytest, OS keychain
- Tests
- 45
- Services
- AI integration
Data, never instructions
Every channel returns a nonce-fenced provenance envelope that marks content as untrusted, strips invisible characters and reports whether anything in it tried to address the agent. There's no code path that skips the wrapper.
Hardened against the original
Compared with the upstream project (Panniantong/agent-reach, MIT), ReachKit pins external tools to exact versions, keeps secrets in the OS keychain, reads the web directly instead of through a third-party proxy, and closes two server-side request forgery gaps.
Keep exploring.
Want something like ReachKit?
Tell us what you're trying to build. You'll hear back within one business day with questions, a suggested approach and a ballpark.