Lex
Repo-local memory and policy context. Frames, recall, Atlas neighborhoods, policy checks, instructions, CLI, MCP, and TypeScript API.
Local-first agent infrastructure
An operating stack for agent work.
SmarterGPT is building the pieces agents need after the model call: durable memory, workspace-native capabilities, emerging authority boundaries, behavioral posture, and deterministic orchestration.
The first application is software development. The same pattern applies to analyst workflows, operations, compliance review, support, and any environment where agents need bounded autonomy instead of a pile of prompts.
Lex public MIT npm latest
AXF public source-available npm latest
LexThority pre-alpha design preview
LexRunner and LexSona controlled-access
Current Stack
The public surface is intentionally technical. Each piece has one job, a clear boundary, and a different maturity level.
AXF
Agent eXoskeleton Framework. Workspace-owned capabilities that tell agents how this repo builds, tests, searches, diagnoses, and operates.
LexThority
Pre-alpha authority seam research. Explore how bounded agent autonomy, effect boundaries, escalation, and receipts should fit together.
LexSona
Behavioral constraint engine. Personas and learned rules become deterministic constraints. It returns constraints; it does not execute.
LexRunner
Flow runner. Fan-out work, run gates, coordinate contained execution, and merge through deterministic plans.
LMV Precepts
LMV is the evidence-aware memory doctrine moving upstream through Lex and AXF. These 12 precepts keep memory, evidence, uncertainty, power, and operator control separate.
Evidence is not truth.
Evidence can support a claim without becoming final authority.
Memory is not evidence.
A Frame can guide inquiry, but evidence closes validation.
Recall is not certainty.
Recovered context remains a candidate until checked against current state.
Uncertainty is part of truth.
Unknown, partial, stale, and low-confidence states should remain visible.
Contradiction is signal.
Conflicting records should be linked and resolved, not overwritten.
Forgetting is a feature.
Retention, demotion, expiry, and supersession keep memory useful.
Assimilation must add distinctiveness.
Promote memory when it changes future behavior or prevents repeated work.
Provenance is dignity.
Claims deserve cited witnesses: files, commands, logs, receipts, commits, URLs, or Frames.
Boundaries are consent.
Records and capabilities should say what they do and do not authorize.
Power requires audit.
Meaningful effects need actor, scope, reason, result, and validation records.
The operator remains sovereign.
Humans remain responsible for destructive, secret-bearing, remote, or unbounded actions.
Failure under discipline is inquiry.
Bounded failures can teach when scope, cause, effect, and limits are recorded.
Why this exists
Models are not the operating system
Frontier models can reason, but production work still needs local memory, repo facts, tool contracts, permission boundaries, receipts, and failure recovery.
MCP exposes tools; it does not define judgment
MCP is an important transport. It does not decide whether an agent should deploy, which credentials are appropriate, or what to do when access fails.
Autonomy needs provenance
Rigid SOP automation turns agents into brittle programs. Unbounded autonomy turns tools into risk. The useful middle is bounded work with receipts.
Available Today
Use Lex in real repos
Install Lex to give agents local memory, policy context, recoverable errors, and durable instructions. It is MIT licensed and published on npm.
Use AXF for workspace capabilities
AXF packages repeatable repo know-how as capabilities. Agents can inspect, run, and promote those capabilities through one contract.
Preview LexThority
LexThority is public as a pre-alpha design preview for the authority seam. The concept is useful, but the boundary is still expected to change.
Beyond Development
The software-development loop is the proving ground because it has clear tools, artifacts, and tests. The architecture is broader than code.
Data analysis
Analysts need dataset memory, query capability declarations, role-scoped credentials, and receipts for decisions made from uncertain data.
Operations
Ops agents need runbooks converted into inspectable capabilities, authority checks before mutation, and escalation when a credential or policy blocks action.
Product and support
PM and support agents need different access than developers. LexSona can express posture; authority boundaries remain explicit; Lex records what happened.