SHADI
Secure Host for Agentic AI Dynamic Instantiation
SHADI is a hardened runtime for agents operating near real credentials, local data, and developer tooling. It enforces launch-time policy through verified identity, gated secret access, OS-level sandboxing, encrypted local memory, and secure transport — so agent autonomy never comes at the cost of a real security boundary.
Why SHADI
Verified Identity
Deterministic human-to-agent derivation with provenance checks, so every session is traceable back to a real, authorized principal.
Secrets Gate
Keychain-backed secrets, with optional 1Password integration, released only to sessions that pass identity verification.
Kernel Sandbox
OS-enforced policy — not prompt intent — with portable profiles and JSON policy support for read, write, and network access.
Encrypted Memory
SQLCipher-backed local state keeps agent memory encrypted at rest, on disk, between runs.
Secure Transport
DID-authenticated, MLS-encrypted messaging between agents via SLIM and A2A.
Bring Your Own Coding Agent
agentbridge lets Claude Code, Codex, GitHub Copilot, Cursor Agent — and any other CLI coding tool, including Gemini CLI, via a generic adapter — hand off context, delegate tasks, and coordinate autonomously with each other over A2A on SLIM. Zero trust throughout: every agent authenticates with its own DID, never a shared secret.
See shadictl in Action
Create a moderator-owned SLIM channel, invite coding agents by DID, and broadcast to the whole group with A2A Collaborate. See the Secure Agent Group Demo for the full walkthrough.
Launch a real command inside SHADI's sandbox, with a secret released only to the verified session. See the CLI Reference for the full flag set.
Inspect and diff the resolved sandbox and secret-delivery policy before a process ever launches. See Operations for day-to-day workflows.
Community
Connect with AGNTCY contributors, join working group meetings, and help shape secure agent runtimes. For more information, see the community page.
Contribute
Help build SHADI by contributing code, reporting bugs, or suggesting enhancements. Pick up a good first issue, review open pull requests, or read the contributing guide to get started.