EdgeFlow Join the waitlist

Run your AI agents on hardware you own.

EdgeFlow deploys the agent you already use onto a Mac mini, a DGX Spark or your own servers, with a local model, a sandbox and an audit trail. Private and governed by default.

Join the waitlist See how it works

The open-source runtime, v0.1, is in development. We are choosing five design partners from the waitlist.

$ edgeflow init ./ops-agent
							✓ agent   OpenClaw, 4 skills, 2 channels
							$ edgeflow deploy --target spark.local
							✓ model   gpt-oss-120b on the local GPU
							✓ sandbox network blocked, 3 hosts allowed
							✓ audit   every tool call logged locally
							✓ memory  facts rechecked every 30 min
							running at https://spark.local:8700
Planned CLI, shown as a design preview.

Agents need your most sensitive data. Governance says it can't leave the building.

A coding agent reads your whole repository. An operations agent reads contracts, tickets and patient records. Every step sends that context to a third-party model, so the most regulated teams opt out.

27%
of organizations have banned generative AI apps.
56%
now run or plan to run production AI inference privately.

Sources: Cisco 2024 Data Privacy Benchmark; Broadcom Private Cloud Outlook 2026.

Your hardware is the boundary.

Everything the agent needs runs inside it. Nothing leaves unless you allow it, and every step is written down where you can read it.

Your Mac mini, Spark, server or VPC

  • Your agent
  • Local model
  • Sandbox
  • Audit trail

The internet

Blocked by default. Allow a host, such as your Git server, and every request to it is logged.

Bring your agent
Claude Code, Codex, OpenClaw, LangGraph or your own code. No rewrite.
Model gateway
An Anthropic- and OpenAI-compatible endpoint served from your GPU.
Agent-aware inference
Caching and routing tuned for long agent loops, not single chats.
Sandbox
Every run is isolated. Outbound network is blocked unless you allow a host.
Governance
Policies, access control and a full audit trail, kept on your hardware.
Self-checking memory
What the agent believes about your systems is rechecked on a schedule and flagged when it goes stale.

Runs on the machines teams are already buying.

Desk-side GPUs, open-weight models like gpt-oss and agents that call any compatible endpoint made private agents practical this year.

  • Mac mini and Mac Studio A quiet desk-side node for one team.
  • NVIDIA DGX Spark Runs 200B-parameter models on a desk.
  • On-prem servers Your racks, your network, air-gapped if required.
  • Your cloud VPC Inside your own account, never ours.

Land with developers. Expand into regulated teams.

Startups

You already run agents on a Mac mini or a Spark. Start on the free tier, then add teammates and nodes.

Finance and health

Contracts, tickets and patient records stay inside the building, with audit trails, access control and residency.

Government

Air-gapped and sovereign by mandate. Enterprise licenses with a support SLA.

Open core. Pay per node when your team grows.

Community

Free open source

For individual developers

  • One node, any agent
  • Local models and sandbox
  • Local audit log

Pro

$99 per node per month

For startups and teams

  • Multi-node fleets
  • Shared agent queue
  • SSO and role-based access
  • Policy templates

Enterprise

Annual license

For banks, hospitals and government

  • Air-gapped installs
  • Audit export to your SIEM
  • Data-residency controls
  • Support SLA

Built by engineers who care about software and hardware.

Our team has studied and worked at

  • University of California
  • University of Pennsylvania
  • Stanford
  • Hanwha
  • Intel
  • LG
  • G42

Run your first agent on your own box.

Tell us which agent you run and on what hardware. Design partners from the waitlist get the runtime first and shape what ships.

Join the waitlist