Ostivane

Runtime security for AI agents. Ostivane checks every tool call an agent is about to make, refuses the dangerous ones before they run (destructive commands, leaked secrets, data sent to drop sites, actions after a prompt injection), lets a person decide on borderline ones, and keeps a signed, hash-chained record your auditors can verify offline.

pip install ostivane                  # SDK, CLI, gateways
pip install "ostivane[web]"           # the hosted cabinet (self-host it)
pip install "ostivane[mcp]"           # MCP gateway and server

See it work in a minute, with nothing to sign up for: ostivane try walks a demo agent into a prompt injection and shows which of its next steps are refused, and why. Add --cloud URL --workspace WS (and P7_API_KEY) to watch it in your cabinet.

from ostivane.recorder import AgentRecorder

rec = AgentRecorder("https://app.example.com", API_KEY, WORKSPACE, agent="support-bot")

@rec.tool("bash", target=lambda cmd: cmd)          # checked before it runs; a refused try is recorded
def bash(cmd): ...

No code change: ostivane soc-llm (OpenAI, Anthropic, Gemini), ostivane soc-mcp / ostivane soc-mcp-http (MCP servers), ostivane soc-gateway (HTTP proxy). Clients for TypeScript and Go are in the source repository, which is not public yet.

Import it as ostivane. Code written for the earlier present-day package keeps working: import present_day gives the same modules, and the present-day command is still installed.

Licensed under Apache-2.0.