Agents & decoys: choose the job
These features solve different problems. Use a framework adapter to let an agent call a tool that uses a real secret. Use a decoy control file to open encrypted data to a chosen fake. Use Honey Mode for fake outputs from incorrect password/control combinations in supported formats.
Generating a fake does not register a monitor. For an alert, add a tripwire and test the monitored path. A decoy is not a working provider credential.
Agent framework adapters
Connect an agent to a tool that uses a secret without returning the raw secret to the model. The adapters resolve the credential inside the tool and return only the fields your tool is designed to return. By default, they reject a result containing the raw secret. This check does not detect every transformed or indirect leak; your access policy, logs and error handling still matter.
# LangChain (Python / TypeScript)
pip install deny-sh-langchain
npm i deny-sh-langchain @langchain/core
# Vercel AI SDK (TypeScript)
npm i deny-sh-vercel-ai ai
# OpenAI Agents SDK (Python / TypeScript)
pip install deny-sh-openai-agents
npm i deny-sh-openai-agents @openai/agents
# Shared framework-agnostic core
npm i deny-sh-integrations-coreExports are identical across adapters: denyVaultTool / deny_vault_tool, createVaultResolver / create_vault_resolver, isNarrowed / is_narrowed, DenyToolError, DenyLeakError. See the integrations pages for working agent loops.