DUBSAR GitHub

How it works

One product: DUBSAR. A Claude Code plugin with a local runtime.

DUBSAR adds decision memory and human governance around Claude Code.

You install the DUBSAR plugin in Claude Code and DUBSAR Desktop locally, run one unified start, and DUBSAR keeps your Mission, constraints, evidence and Human Gates coherent across sessions.

Join the private beta Read public docs

Private beta. Built for developers working with Claude Code.

Problem

Speed does not automatically create coherence.

Claude Code is fast. It generates, explains, fixes and reviews in minutes.

But on long-running projects, decisions accumulate and context changes.

Constraints get forgotten between sessions. A model can produce a reasonable answer and still move the project in the wrong direction.

The question is no longer only whether AI can write the code. It is whether the project can remember what was decided, what was locked, and what evidence supports the next change.

One product

DUBSAR is one product with local components.

  1. 01 · Plugin

    The DUBSAR plugin for Claude Code

    A thin plugin. It carries no proprietary logic — it reflects the Mission, memory, evidence and Human Gates inside Claude Code.

  2. 02 · Desktop

    DUBSAR Desktop — the local runtime

    A required local component that supports the plugin (local Bridge + cockpit). Not a separate product, not a second offer.

  3. 03 · Private

    Private Backend & Core

    The governance engine runs privately. Nothing proprietary is distributed in the plugin.

What DUBSAR keeps around the work

DUBSAR does not replace Claude Code.

It gives long-running work a structure — a Mission, locked constraints, lots, contracts, evidence and Human Gates — and a replayable trace.

Goal

Make long AI-assisted projects easier to trust, audit, and replay.

Governed workflow

  1. Mission
  2. Locked constraints
  3. Lot
  4. Contract
  5. Execution
  6. Evidence
  7. Human Gate
  8. Replay

What DUBSAR does

What it puts around the work.

  1. 01 · Memory

    Persists a Mission beyond one session

    The project keeps its Mission, decisions and locked constraints across restarts and context compaction.

  2. 02 · Structure

    Structures progress by lots and contracts

    Work advances through lots and contracts instead of open-ended prompts.

  3. 03 · Evidence

    Links execution, evidence and audit

    Reports, proofs and audit stay connected to the change they belong to — read-only, reviewable.

  4. 04 · Replay

    Keeps a replayable path

    The decision trace can be replayed, so a long project can be picked back up with its context intact.

  5. 05 · Human

    Reflects where a human must decide

    DUBSAR does not let an agent validate itself. Human Gates stay explicit.

Why not just use an agent?

A proposal is not a governed decision.

Claude Code is useful. But an agent’s proposal is not the same as a governed decision.

A long project needs more than generated code. It needs:

  • Decision memory
  • Locked constraints
  • Contracts
  • Evidence
  • Replay
  • Human approval

Agents propose.
DUBSAR checks.
Humans decide.

Install & run

From install to a project with memory.

  1. Get private beta access

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

  2. Install DUBSAR Desktop

    The local runtime that supports the plugin.

  3. Install the DUBSAR plugin in Claude Code

    The plugin appears inside your Claude Code workflow.

  4. Run the unified start

    DUBSAR recognizes the workspace and resumes or creates the Mission.

  5. Work with memory and a human gate

    The local cockpit reflects lots, contracts, evidence and Human Gates while Claude Code codes.

Governed workflow demo

Mission to replay, with a human in the loop.

Static, non-executable demonstration of the governed workflow: Intent, Contract, Instruction Packet, Diff Audit, Human Decision, Replay. No agent is connected.
Governed workflow — demo STATIC DEMONON-EXECUTABLEREAD-ONLY
  1. Intent
  2. Contract
  3. Instruction Packet
  4. Diff Audit
  5. Human Decision
  6. Replay

01 · Intent

Intent detected
Refactor auth module.

02 · Contract

Allowed paths
src/auth/**
Forbidden paths
.env, billing/**, infra/**
Expected tests
auth unit tests
Human approval
required before merge

03 · Instruction Packet

Agent target
Claude Code
Mode
dry instruction only
Execution
disabled until explicit human GO

04 · Diff Audit

Modified files
4
Out of scope
0
Forbidden changes
0
Tests
pending passed failed

Demonstration states, not a live result.

05 · Human Decision

Approve Reject Request changes

display-only — the real decision stays with the human, outside this demo

06 · Replay

intent#a41f contract#9c02 packet#d812 audit#7af0 decision#c31e

Decision path preserved.

Static demonstration. No agent is connected to this website.

Built for

  • Developers using Claude Code on real projects.
  • Indie hackers building serious products with AI assistance.
  • Small teams that want AI speed without losing project memory.
  • Technical founders who need proof, not just chat history.
  • Anyone who has felt a project drift after too many AI-assisted changes.

What DUBSAR is not

  • DUBSAR is not an autonomous coding agent.
  • It is not a replacement for Claude Code or for developers.
  • It is not a marketplace of agents.
  • It does not make models more intelligent.
  • It does not remove human responsibility.
  • It is decision memory and human governance around Claude Code — not commercial-ready.

Private beta

DUBSAR is in private beta.

Access is selective and reviewed manually, focused on real long-running Claude Code projects. Windows first; macOS and Linux follow.

Request beta access

FAQ

Questions

Is DUBSAR another AI coding agent?

No. DUBSAR is not the agent. It is decision memory and human governance around Claude Code.

Does DUBSAR write code by itself?

No. Claude Code writes the code. DUBSAR preserves decisions, constraints and evidence, and keeps a human approving.

Do I need DUBSAR Desktop?

Yes. DUBSAR Desktop is the local runtime the plugin needs — a required local component, not a separate product or a second offer.

Which tools does it work with?

The private beta is built for Claude Code. Codex, Cursor and broader multi-agent orchestration are future direction, not beta scope.

Is the engine open source?

The public documentation describes the doctrine and principles. The private governance engine stays private and is not distributed in the plugin.

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

DUBSAR keeps long-running AI-assisted development bounded, reviewable, and governed.

Request beta access Read public docs