Trust by construction

Show your work. Every claim, every run.

Kermit is safe to point at real systems because it can't change them. Read-only inputs, draft-only outputs, and an audit trail rich enough to argue with.

Source health & evidence

If a source is down, you'll know — before it shapes a decision.

A preflight runs against every connector at the top of each pipeline. Healthy, partial, missing and stale states are surfaced explicitly. The dashboard never silently drops a source.

Per-connector status: healthy / partial / stale / missing
Every priority card carries a [source] link to the underlying record
Run logs record what was reachable and what wasn't

Source health panel with status pills for each connector and an evidence list of source-linked items underneath.

Human-approved memory

The only memory Kermit keeps is the memory you've approved.

Patterns Kermit notices are proposed, not adopted. You review, edit, accept, or discard each candidate before it influences another run. Memory changes are recorded like any other write.

Memory candidates queued, never written silently
Diff view: what changed, why, and from which evidence
Roll back any accepted memory with one tap

Memory review screen with candidate memories, supporting evidence, and accept / edit / discard controls.

Six hard rules

Enforced by construction, not by policy.

Draft-only

Never sends, posts, or submits. Output is always a draft for you.

Read-only external

No record in any connected system is ever modified.

No secrets in memory

Credentials and sensitive data are never persisted.

Epistemic tagging

Every claim is labeled FACT / INFERENCE / ASSUMPTION / REVIEW.

Fail-loud

Missing sources are flagged by preflight, never silently skipped.

Work / personal split

A strict boundary; no personal data in work artifacts.

Bring Kermit into your stack.

We onboard new workspaces with a guided source-health pass and a sandboxed first run. No writes until you say so.

mykermit.aiA Blank Page project