Soba
Soba is an LLM layer for agentic applications that lets developers route AI runs to the Claude or ChatGPT plans their users already pay for, eliminating the inference costs of free tiers. Developers integrate through a single OpenAI-compatible API endpoint, changing only the base URL and adding a user identifier. The platform handles routing, permissions, cost classification, and metering, with billing compiled into the developer's own Stripe account.
- Developer Tools
- Artificial Intelligence
- Software Only
Funding
Founders
Founder details are not available yet.
Product
Problem
Every free user of an agentic AI application costs the developer money because every run is inference the developer pays for. This forces developers to ration free tiers with credits, paywalls, and caps, which often end before users have built the habit of using the product. A longer, more generous trial is the obvious fix but one most teams cannot afford, as its cost grows with how generous it is and how many people take it.
Solution
Soba is an LLM layer for agentic applications that lets users run AI features on the Claude or ChatGPT plan they already pay for. A user connects their own AI account—via OAuth or a paired machine—and Soba routes each run to that user's own computer and subscription, meaning those runs cost the developer nothing to serve. The platform enforces a strict invariant: a run executes on the end user's own machine, under their own login, for their own account, with attribution enforced in code rather than documented. Developers integrate through a single OpenAI-compatible API endpoint, where the only changes are the base URL and a user identifier, so the call site never needs to know where a run executed. Soba handles routing, permissions (intersection of what the machine owner allows and what the app requests), cost classification (user hardware, user subscription, open-weights, or frontier), and metering in integer micros. It also provides billing infrastructure, compiling pricing plans into real prices in the developer's own Stripe account, so developers set their own rates and keep the margin.
Target Audience
Primary customers are developers and product teams building agentic AI applications who want to offer generous free tiers and trials without absorbing the inference costs. This includes startups, SaaS companies, and enterprises deploying AI-powered features that want to convert users who already pay for Claude or ChatGPT plans.
Features
- OpenAI-compatible API endpoint: developers change only the base URL (https://api.soba.so/v1) and pass a user identifier, with no branching or flags for connected vs. unconnected users
- Worker that dials out from the user's machine over a single WebSocket, reaching machines behind home routers without public addresses or port forwarding, and surviving sleep, wake, and network changes
- Permission intersection model: a run gets the overlap between what the machine's owner allows and what the app requests, a resolution that can only ever narrow
- Cost classes ("user-hardware", "user-subscription", "open-weights", "frontier") that are asked rather than assumed, with the worker querying `claude auth status` locally to determine who actually pays
- Attribution binding: pairing tokens are bound to the one person who approved them, preventing pooling or shared account misuse
- Settings isolation: runs arriving over the network do not inherit the machine owner's CLAUDE.md, hooks, skills, or plugins
- Dashboard with the lead metric of "inference you didn't pay for this month," auditable against the developer's own provider bill, plus per-user usage, connected machines, spend ceilings, and failures
- Envelope dials for trials: duration, volume, and feature scope, allowing free periods that last until the user's habit is formed without inference costs landing on the developer