Local, Open Source AI
Inference Exchange
Buy by day, sell by night.
Portedges is the first AI Inference Exchange at the edge.
By night, list your high-performance hardware when it sits idle and sell local, open source AI Inference to the market. By day, buy access to open-source AI inference capacity at the edge and scale your AI routines (skills, loops and batches) without expensive frontier model API costs. Post asks, place bids, experience market's price formation and tokenomics in real time.
The price, per 1M output tokens.
What's in It for Me
If you build with AI agents, you have probably met the local glass ceiling. Your machine handles one modest agent at a time, and the moment you want several skills firing on a schedule, or a long self-correcting loop running overnight, you need far more memory and compute than you own. You are hardware-constrained, not idea-constrained: the ambition exceeds the silicon.
The usual escape hatch is a metered frontier API, and for repetitive, heavily batched, non-frontier background work it is a cost trap : you burn frontier-model budget on tasks that do not deserve it. Between "my Mac chokes" and "I rent expensive cloud" there has been no affordable, on-demand, trustworthy middle option. Portedges is that middle option: capacity for routine agentic work, reached through an OpenAI-compatible endpoint your existing setup already understands, so you scale your agentic ambitions without buying new hardware.
And if you are on the other side of the market — if you own a high-performance Apple Silicon Mac that sits dark eight to ten hours a night — that is paid-for capacity doing nothing. Portedges is where you list, ask for a price to access your Apple Silicon inference capacity and when a bid matches your ask you let it work while you sleep. In this first release only Apple Silicon Macintosh and the oMLX inference engine are supported. Nvidia GPUs and Ollama will be available soon.
What is Portedges
Portedges is an inference exchange. On one side are developers whose agentic workloads have outgrown their own machines; on the other, people whose powerful, already paid-for hardware sits idle overnight. Portedges is the market between them: it lists what is available, forms a price, brokers the trust, and settles the value.
Connecting machines you already own is essentially a solved problem. The hard and valuable part — the reason Portedges exists — is enabling exchange between people who do not know each other. Trust runs both ways: the provider has to trust a stranger's workload on their machine, and the buyer has to trust a stranger's machine with their data and workflows. Both directions are the market's job, which is why the market is the core of the product and not a feature of it.
Three familiar patterns explain the shape. The interface behaves like OpenRouter: one endpoint standing in front of many backends, so your agent's configuration barely changes. The marketplace behaves like Uber or Airbnb: it turns an idle liability into an income-generating asset, with supply and demand drawn from overlapping populations. And the mechanism behaves like a spot exchange, because idle compute is perishable — capacity going unused tonight is worth nothing tomorrow.
The early market is deliberately prosumer-first, and the same person is often on both sides: buying capacity by day, providing it by night. Portedges follows the broker pattern — it takes a share of the transactions it enables; it does not own the hardware and it does not build the models.
Why Now
Three windows are opening at the same time. Agentic workflows have just gone mainstream: the shift from prompting to loop engineering — designing loops that prompt agents on a schedule, with no human in the loop — is recent and accelerating, and it manufactures exactly the bursty, concurrent, long-running workload that overwhelms local hardware. The behaviour creating the pain is emerging right now.
At the same time, capable open-source models now run locally. Non-frontier models that are genuinely good enough for day-to-day chores fit on consumer Apple Silicon, which makes "good enough and cheap" a real alternative to a frontier API for routine work. And the installed base of high-memory Apple Silicon is large and still growing, so the idle supply is real, substantial, and expands with every hardware cycle.
There is a fourth reason the timing works, and it is structural rather than technological: demand and supply are naturally offset around the globe. A productive afternoon in Casablanca is the middle of the night in Montreal, and twelve hours later the roles reverse. The geographic spread of the community is the supply schedule — which is a structural reason a market like this can clear at all.