Project knowledge stays local
Project Brain, project bindings and confirmed project knowledge are stored locally. Livariant does not automatically upload them to a Livariant cloud service.
LivariantWindows Desktop Preview0.1.0-rc.29Local-first
Local control for Codex, Claude Code, Gemini CLI and other coding agents
The control layer for AI coding agents.
Livariant sits between your coding agent and your project. It preserves reliable context, keeps consequential actions behind explicit authority boundaries, and verifies results against evidence instead of trusting claims at face value.
Local desktop and CLI applicationconnects through MCPto your coding agent
AI OUTPUT / UNTRUSTEDPROJECT TRUTH / DURABLE
Observationnot project truth yetContextprovenance stays visibleAuthorityexplicit decisionVerificationevidence-backed state
Your agent keeps working in the interface you already use. Livariant keeps context, authority and evidence separate in the background — across sessions, providers and projects.
01
What Livariant protects
Four safeguards for agentic software development.
Without LivariantAgent risk
With LivariantProject control
×
Context gets lostImportant decisions disappear between chats, models and providers.
✓
Project knowledge stays durableConfirmed context remains local and belongs to the project, not to one conversation.
×
Capability becomes permissionAn agent may be technically able to do more than you intended to authorize.
✓
Authority stays explicitWhat a system can do and what it is allowed to do remain separate.
×
“Done” remains a claimA convincing summary is not proof that the requested result exists.
✓
Evidence makes results verifiableClaims are evaluated as supported, unproven or contradicted.
×
Wrong assumptions become permanentA mistaken interpretation can silently become part of the project state.
✓
Recovery is part of the modelConsequential changes remain bounded, traceable and recoverable where supported.
Livariant does not replace your coding agent. It gives the agent a controlled working environment that survives individual chats, models and provider sessions.
02
In the normal workflow
You direct the agent. Livariant protects the path.
- 01Inspect
Read
Capture the existing state
- 02Project Brain
Understand
Reconstruct project context
- 03Provider Handoff
Propose
Prepare the change and context
- 04Authority
Authorize
Grant consequential permission separately
- 05Recovery
Change
Perform the bounded mutation
- 06Verification Trace
Verify
Check the result against evidence
03
Verification Trace
“Done” becomes verifiable.
Livariant compares each requirement with the agent’s claim and the evidence that actually exists. That makes it immediately visible what is supported — and what is not.
RequirementClaimEvidence status
Email sign-inimplementedSUPPORTED
Password resetimplementedUNPROVEN
Rate LimitingimplementedCONTRADICTED
SUPPORTEDThe evidence supports the claim
UNPROVENThe available evidence is insufficient
CONTRADICTEDThe evidence conflicts with the claim
VERIFIED ≠ ACCEPTED The final decision remains yours.
04
Local-first
The control layer runs where your project lives.
Connections are deliberate
GitHub and external AI providers are only used through their documented connection paths. Connecting a provider does not grant mutation, merge or release authority.
No silent product telemetry
The current product does not include Livariant usage analytics, crash reporting or advertising identifiers. Desktop update checks are initiated by the user.
05
Available today
Published preview
Desktop Preview 0.1.0-rc.29
Windows · immutable qualified source identity · not a Stable release
Available separately
Core / CLI Preview
Versioned independently · repository state and CI alone are not release evidence
Try Livariant