Docs Frameworks

LlamaIndex Agents

Status: preview. The adapter is planned, not shipped. The integration shape below is the survey’s working assumption; it has not yet been source-verified the way the four shipped adapters were, so treat it as direction, not a runnable path. This page flips to verified when the adapter lands in kiff-guard and passes the conformance suite.

Expected shape: inverted-control. LlamaIndex agents are expected to run the tool themselves with a vote-style hook, so the adapter would use Guard.observe() / Guard.decide_only() rather than a run callback — the same shape as Hermes and the OpenAI Agents SDK. The exact seam is confirmed during the adapter PR’s source-verification step.

What “done” will look like

When the adapter ships, this page carries the same spine as every verified framework:

  • Install — the guard core plus the LlamaIndex extra.
  • Observe — attach the guard in observe mode, run the agent, read the audit trail. Zero config, no account.
  • Activate — derive a starter domain with export_yaml and activate it in the dashboard.
  • Enforce — point the guard at a tenant, switch to enforce, and KIFF decides before each tool runs.

The model is identical across frameworks; only the one-line attach and the seam differ. Read how the guard works for the shared model.

Want this sooner?

Adapters are community-friendly by design: each is ~15 lines over the agnostic core, and “done” means passing the conformance suite. Track or contribute at kiffhq/kiff-guard.

← Quickstart kiff-guard on GitHub ↗