AI should do useful work, remember why, and show its evidence.
I'm Brandon Ducar, building DreamNet: connected tools and contracts for AI work that survives the chat. The aim is a shared operating system for digital and physical businesses. The public projects tackle concrete pieces: persistent context, bounded work, source-linked knowledge, and results you can inspect.
Your AI says "fixed." BREAKSHIFT says: prove it.
Inspect a component, challenge a repair plan, and replay the evidence in a simulated robotic workcell. Change the assumptions and stale passes are withdrawn. Simulation is not physical repair clearance.
Try the open demo | Watch the corrected v2.2 demo | Inspect the source
The workbench opens without a login. Fresh paid model analysis remains operator-restricted; retained model outputs are labeled as recorded.
DreamNet's design draws from natural systems and human engineering, not one giant agent or a literal artificial human. The direction is simple: keep what works, improve its connections, test the result, and preserve what we learn. That is an engineering goal, not a claim of autonomous recursive self-improvement or a verified active fleet. Each repository states what it actually implements.
Explore the project directory | Ecosystem map | DreamNet
Independent products. Optional connections. Private operations. Start with one project's documented setup; you do not need access to DreamNet's private monorepo. Each product keeps its own requirements, license, data boundaries and release lifecycle. Integrations use explicit interfaces, not shared credentials.
Darkhalo's actual public interface. Its built-in sample is explicitly synthetic; the screenshot is not a receipt for real-world work.
Try one reproducible result: Inspect a repository and keep its evidence. Cerberus inspection feeds Memory Weaver's source-linked context export. Runs locally without dependencies or API keys; includes tampering, changed-input and blocked-install cases. Synthetic inputs, actual execution, unsigned receipts.
| I need to... | Try | What it provides |
|---|---|---|
| Preserve and resume bounded agent work | DreamLoops | Versioned operating loops, capability Capsules, execution records |
| Inspect a repo before running it | Cerberus | Offline static inspection and digest-bound findings, not a safety guarantee |
| Review agent-run context | Darkhalo | Local-first inspection and review capsules, no automatic authority |
| Prepare source-linked knowledge | Memory Weaver | Portable context with provenance |
| Hand work to someone else | Warper Keeper | Source-first Trapper context bundles |
| Package evidence for a build | Proof Drop | Local evidence-file hashing and exported receipts |
The independently operated ZOL project merged a DreamLoops persistent-agent integration and a Bonfire memory adapter. The pull requests show the contributions. They are not a claim about ZOL's current runtime, an endorsement, or DreamNet ownership.
- Public Core: shared agent contracts.
- ToolGym: agent workouts; practice is not independent qualification.
- LaunchCast: repository-to-storyboard and local video rendering; uploads are not implemented.
- Dreamnetiopi: merchant-operations rehearsal, not production telemetry.
- Public project directory: a dated catalog of apps, libraries, research, prototypes, and attributed forks, not a live inventory.
| What brings you here? | Next step |
|---|---|
| I need help or found a bug | Open a product help request |
| I want to connect two tools | Propose an integration |
| A link, screenshot or description is wrong | Report a directory correction |
| I have a broader question | Join the public-core discussion |
| I want to contribute code or docs | Read the contribution guide |
Not sure which repository owns a problem? Start here with the product name and a small, non-sensitive example. We can identify the right repository before you invest in a patch. These are public conversations, not private support tickets.
Support and privacy boundaries
DreamNet is early. Each repo has its own implementation boundaries and license. A receipt is not proof that every claim inside it is true, and a capability is not permission to act. Private credentials, customer data, and operating context do not belong in public examples.
Public overview revision: 1.0.0, November 4, 2026. This is a documentation revision, not a synchronized release of every linked product.