Repository-native
Intent, state, decisions, and technical intelligence live with the project instead of disappearing inside chat history.
Keep agent work
bounded, durable, and reviewable.
Sculptus is a repository-native control plane for AI software agents. It keeps project intent, current state, task scope, approvals, and validation evidence in durable repository contracts.
Founder-operated · actively used across private repositories
ACTIVE TASK
✓
Read canonical stateRepository is the source of truth
✓
Bind task scopeAdjacent work stays parked
↻
Execute with bounded toolsMCP access follows explicit contracts
4
Validate evidenceTests, diff, and reproducible checks
5
Human approvalNo silent consequential action
Durable stateProject truth survives chat context.
Human gateCritical decisions stay under owner control.
Bounded agent work
Each task carries explicit scope, acceptance criteria, stop conditions, and a place for adjacent findings.
Controlled access
MCP exposes only the project context and operations the agent needs, rather than giving every session unlimited reach.
Evidence before approval
Tests, Git evidence, deterministic checks, and review remain part of the workflow before consequential changes are accepted.
HOW SCULPTUS WORKS
A control layer between intent and execution.
Sculptus does not replace the coding or reasoning agent. It gives agents a durable operating contract and keeps the project authoritative.
1Human intentDirection, constraints, approvals.
2Repository truthProject, state, decisions, intelligence.
3Bounded taskScope, gates, expected evidence.
4Agent executionCoding, research, review, or analysis.
5ValidationTests, diff, reproducible evidence.
6Human approvalConsequential next steps stay explicit.
PROJECT MEMORY
Small files. Clear responsibilities.
project.mdstable intent
state.mdone active objective
decisions.mdappend-only decisions
ideas.mdparked or selected
project-intelligence.mdarchitecture & invariants
MCP
Explicit context, bounded tools.
Every project is named explicitly. Read and write surfaces are narrow, allowlisted, and separated by role.
projectNamerequired
scopecurrent-task
sideEffectsowner-approved
VALIDATION
Agent output is not authority.
Implementation claims are checked against code, tests, Git state, and task-specific deterministic gates before closure.
✓ target checks
✓ full regression when required
✓ Git evidence
✓ owner review
DESIGN PRINCIPLES
Built for long-running AI-assisted projects.
01
Repository over memory
Chat history and model memory are context, not canonical project truth.
02
One active objective
Adjacent ideas are parked instead of silently expanding scope.
03
Model-independent control
The workflow is designed around contracts and evidence, not dependence on one agent provider.
04
Human authority
Important product decisions, external effects, merges, and deployments remain explicit owner choices.
CURRENT STAGE
Private by default. Proven through direct use.
Sculptus is actively used across multiple private AI-assisted repositories spanning software development, research, and evidence-heavy workflows. Claude was integrated into the earlier Sculptus workflow. In v2, execution moved to another coding-agent stack while the underlying control layer was rebuilt around provider-independent project state, bounded access, and explicit approval gates. The codebase remains private.
CONTACT