DUBSAR GitHub

Private beta

Keep long-running Claude Code projects coherent.

DUBSAR is decision memory and human governance for AI coding.

Claude Code builds fast. On long projects, the decisions, constraints and evidence behind the code get lost between sessions. DUBSAR keeps them — in the project, not the chat — and keeps a human in control.

Join the private beta See how it works

Built for developers using Claude Code on long-running projects.

The problem

Speed does not create coherence.

  1. 01 · Speed

    Code moves faster than decisions

    Claude Code can generate and edit quickly, but the reason behind a change gets lost.

  2. 02 · Memory

    Context is lost between sessions

    Decisions live in chats and memory that reset on restart or context compaction — not in the project.

  3. 03 · Locked rules

    Locked constraints drift

    A rule that was clear last week can silently disappear from the next prompt.

  4. 04 · Merge risk

    Diffs look correct but move the project wrong

    A change can be technically valid and still break what the project already decided.

What DUBSAR does

Decision memory and human governance around Claude Code — without replacing it.

DUBSAR is not another coding agent. It sits around Claude Code and keeps what was decided, locked, proven, and waiting for human approval — so a long project stays coherent across sessions.

Claude Code builds.
DUBSAR preserves decisions, constraints and evidence.
Humans decide.

How it works

From install to a project with memory, in five steps.

  1. Get private beta access

    Request a beta access key. Activation stays local and human-approved.

  2. Install DUBSAR Desktop

    The local runtime for the plugin — a required local component, not a separate product.

  3. Install the DUBSAR plugin in Claude Code

    Thin plugin: it carries no proprietary logic, it reflects memory and governance.

  4. Run the unified start

    One command. DUBSAR recognizes your workspace and resumes or creates the Mission.

  5. Work with memory and a human in the loop

    The local cockpit reflects the active lot, contracts, evidence and any Human Gates. Claude Code keeps coding; you stay the authority.

Beta capabilities

What is available in the private beta.

  • Claude Code plugin
  • Local DUBSAR Desktop runtime
  • Unified start command
  • Beta activation & access
  • Workspace recognition
  • Persistent Mission
  • Session continuity after restart/compaction
  • Read-only surfaces: lots, contracts, executions, evidence
  • Controlled local journal / trace
  • Local cockpit
  • Human Gate reflected
  • Diagnostic commands

Human control

An agent proposes. A human decides.

DUBSAR never lets an agent validate its own changes. Approval points stay explicit and human — a decision, not an automatic seal.

Agents propose.
DUBSAR checks.
Humans decide.

Human checkpoint

Read-only

The cockpit reflects the active lot, contracts, evidence and any Human Gates — so the human sees what changed and what remains unverified before anything moves forward.

Human approval required before movement.

Trust architecture

A thin plugin. A local runtime. A private core.

The plugin ships no proprietary logic. It reflects memory, evidence and governance from a private engine that runs locally and privately.

  1. Claude Code
  2. DUBSAR plugin
  3. Local Desktop / Bridge
  4. Private Backend
  5. Private Core

What DUBSAR is not

  • Not a replacement for Claude Code.
  • It does not generate the code itself.
  • It is not an AI model.
  • It does not let an agent validate its own changes.
  • It does not decide for you.
  • It does not promise compliance or bug-free code.
  • It is not commercial-ready.

DUBSAR Audit

A professional review for teams already building with coding agents.

When project state is spread across GitHub, tickets, documentation and agent sessions, DUBSAR Audit establishes what was decided, what is supported by evidence, what conflicts, and what still requires human authority.

Bounded mandate.
Read-only collection by default.
Human-validated conclusions.

Explore DUBSAR Audit

Private beta

Future direction

Where DUBSAR is heading — not beta scope.

  • FutureFull multi-session orchestration
  • FutureAutomated parallel multi-agent management
  • FutureCodex & Cursor integrations
  • FutureEnterprise / on-prem

These are direction, not what the private beta does today.

Claude Code can move fast.
Your project still needs memory.

DUBSAR is not commercial-ready. Private beta. Public surfaces only; the private core is not distributed.