The volunteer compute network for AI safety research.

Sentinel runs as the first tenant. The SDK is open; bring your own experiment. AuspexAI is part of the preneurial[works] family — the operational layer where the methodology meets shipped infrastructure.

We are at Phase 0. The worker, SDK, and coordinator do not yet ship. This page exists so you can find us, follow along, and understand what we are building.

What this is

Distributed compute purpose-built for the workload BOINC was not built for.

Long-running multi-agent simulations, drift studies, adversarial-robustness tests — stateful, multi-step LLM and agent research that runs at scales no individual lab budget reaches.

The platform is open infrastructure. Sentinel — the multi-agent LLM behavioral drift research program — runs as the first tenant. Its tenant package auspexai/sentinel-tenant doubles as the reference example for researchers building their own tenants. AuspexAI's value scales with distinct research projects, not more copies of Sentinel. If you have an experiment that fits, we want yours.

What this is not

Three things to clear up before anything else.

No tokens. No blockchain. No on-chain payment. No plans to add any. Volunteers donate; researchers run experiments. That is the whole transaction.
Not a BOINC replacement. BOINC is a 20-year institution running millions of devices for projects we admire. We are complementary — different workload, different engineering era. We learn from their longevity.
Not source-available. AuspexAI is open source under AGPL-3.0 — OSI-approved, network-copyleft. You can fork it, run it, and modify it. Modifications, including those served over a network, must be released to their users under AGPL.
For volunteers

Two participation paths, both first-class.

Pick whichever fits how you want to contribute. Run experiments yourself, or donate compute to others’ research — both stay supported.

Local Available now

Run Sentinel standalone

Download Sentinel, run it on your own hardware, keep your data local. No network, no signup, no telemetry. Full local control.

Sentinel on GitHub Sentinel research →
Networked Phase 1

Join the AuspexAI network

Donate compute to a shared pool. Run a sandboxed worker that contributes to whatever experiments are active — Sentinel today, more tenants over time. Designed for:

  • Sandboxed by default. Resource caps, idle-only mode, no host filesystem access.
  • Anonymous T0 is fully featured. No signup, no email, no PII collected.
  • Signed contribution receipts. Optional public listing on per-tenant contributors page. Tenants commit to acknowledging volunteer compute in resulting papers and preprints. No leaderboards, no scores, no badges.
  • No tokens, no compensation. Donate-only. Recognition welcomed; payment out of scope.

When the worker ships, this page will link to the installer.

For researchers

Bring your own experiment.

AuspexAI is research infrastructure for the field, not a single-purpose Sentinel host. The platform handles distribution, trust, replication, signing, and analysis. You handle experiment design.

  • Volunteer compute is free. Run experiments at scales your budget cannot reach.
  • Trust, replication, and signed audit log are built-in. Your results are defensible.
  • Tenant SDK with active maintenance. Not a coordinator you have to keep alive yourself.
  • Reproducible signed releases. Methods sections write themselves.
  • Sentinel-tenant is the worked reference example. Read its source to see what a tenant looks like.
Phase 0 status. The SDK is not yet shipped. When it is, this page will link to it. If your work would benefit from the platform sooner, reach out — we'd rather hear from you early than late.
Where we are

Phase 0 — Foundation.

Domain, GitHub org, governance, license, code of conduct, research ethics policy, threat model, and positioning are landing now. Code begins in Phase 1.

Follow on GitHub →
Governance & policies

The commitments that come before the code.

Governance

Roles, decisions, conflicts of interest

Volunteer / Platform Contributor / Researcher / Approver / Maintainer. Lazy consensus, public RFCs, documented conflict-of-interest expectations.
Code of Conduct

Community standards and escalation

Contributor Covenant 2.1 with an AuspexAI-specific Enforcement section, including the escalation pathway for complaints involving the Maintainer.
Contributing

DCO sign-off, how to participate

Developer Certificate of Origin (DCO), no CLA. The barrier to a first PR is one git commit -s.
Research Ethics

What experiments are accepted

Dual-use distinction: studying harmful agent behavior to detect and prevent it is permitted with documented containment; producing harmful artifacts for non-research purposes is not.
Contact

Reach out.

[email protected]

Security disclosures: [email protected] — please disclose responsibly.