One request. The right AI for every step.

Describe the outcome, not the model

Type what you want finished. AI Hub sorts it into one of 14 task types, from writing and long-document summary to code, code review, complex reasoning, data extraction, image analysis, image and video generation, research, voice, translation, and embeddings. Classification happens inside AI Hub, so your request is not sent to a provider just to choose a provider. When the classifier is unsure it asks you to confirm instead of guessing.

Capability first, then your priorities

Models that cannot do the job are removed before any scoring: wrong modality, not enough context, a provider your policy forbids, unknown pricing, or a cost over your cap. Only the survivors are ranked, and five modes (Quality, Balanced, Economy, Fast, Conserve) decide what that ranking favours. A cheap wrong answer is still expensive, so accuracy is the first gate. The route AI Hub would pick with every account connected is always shown separately from the best route you can run right now, so a missing key never rewrites the truth and an unconnected model is never presented as runnable.

Compound work becomes a plan you approve once

A request with several deliverables is split into a graph of steps, each with its own capability needs, model, and cost estimate; independent steps can run in parallel up to a limit your policy sets. You see the full graph and approve it one time before any provider is called. Managed workflows survive interruptions, can be resumed, and can be audited later, with minimized usage records and encrypted content for durable tasks.

Your provider keys never leave the vault

API keys are stored encrypted and used by a broker that makes the call for you; the key is never returned to your browser or phone. Hold several accounts per provider, each with a routing priority, a monthly budget, allowed modes, and an environment label. A circuit breaker pulls a failing account out of service and lets it back in gradually, and you can revoke any credential at any time with a recorded reason. Deleting your account revokes every credential AI Hub held for you first, and the deletion is refused if that cleanup fails.

Use the ChatGPT or Claude plan you already pay for

A small Windows companion pairs your official Codex or Claude Code sign-in to your workspace with a one-time code that expires in ten minutes. Access is a two-minute lease renewed by a heartbeat only while the companion stays open; close it and the lease fails closed. Exactly five task types can run this way: write or rewrite, complex reasoning, translate, generate code, and review code. AI Hub never handles the consumer password, cookies, or tokens, and your subscription allowance is never dressed up as API dollars.

Nine providers, connected directly

OpenAI, Anthropic, Google, Moonshot/Kimi, DeepSeek, Black Forest Labs, Runway, xAI, and Roboflow, 26 models in all, every entry carrying its own official source link and verification date. There is no aggregator in the middle: your request goes straight to the provider you connected. If the catalog's evidence ages past the freshness limit you set, routing stops before a model is chosen rather than deciding on facts that have gone out of date.

A signed receipt for every run, with honest cost labels

Each execution returns a signed receipt with the provider's own usage figures. Cost is labelled Exact when the provider reported it, Estimate when it was calculated before the run, and Unknown when the provider gave nothing back. A missing cost stays unreconciled and cannot be coerced into a numeric zero, so a run is never made to look free. The usage view deliberately excludes prompts, outputs, and credentials.

Spend caps that refuse before spending

Set a monthly organization budget, a per-workflow ceiling, and per-account budgets, with a warning percentage and a hard block at the limit. A request that would exceed a cap is refused before the provider is called. Failed calls never book their estimate as real spend. Long outputs are continued in bounded segments with the worst-case cost reserved first, and an incomplete result is refused rather than passed off as done.

Policy with teeth, for the organization and each project

Five roles (Member, Organization Admin, Operator, Auditor, Workflow Approver) keep writing policy, auditing it, and approving work in separate hands. Policy covers which providers, tasks, and modes are allowed, budgets, retention, and high-risk approval, and it is enforced by AI Hub itself so no client can loosen it. Projects bind approved accounts, connectors, budgets, privacy defaults, and members for one client or team; every project value is intersected or capped against the organization ceiling, so tightening is the only direction available. Every revision needs a reason and lands in an audit trail.

Learn from what actually worked

After a run, rate it one to five, mark it accepted or not, flag rework, and pick a reason from a fixed list. That feedback binds to the execution receipt, stores no free text, and becomes routing evidence for that task family. An administrator can also switch on a bounded challenger that sends a small slice of work to one already-eligible alternative under its own spend cap; it is off unless someone turns it on, and it never calls two providers just to compare their answers.

Privacy you can check yourself

Usage and feedback records omit raw prompts. Durable consumer tasks, active workflow content and selected artifacts can be stored encrypted. Deletion of one account or context store is not proof that all backups, billing records or copies held by a provider were erased. An export carries a checksum you can verify with one command. The audit trail is hash-chained and anchored at every start, on every deletion, and on a six-hour sweep, so edits, mid-chain deletions, and truncation are detectable.

One account on web, Android, and Windows

Sign in once with the same username and password on the installable web app, the genuine native Android app, and the Windows companion. Name your device sessions, sign out any of them remotely or all others at once, rotate your password, and keep ten one-time recovery codes. Your security history lists events only; passwords, tokens, prompts, and outputs never appear in it. On Android, screenshots of the app are blocked.

Beyond text: images, files, code, and cited research

Generate images, proven live through OpenAI, Google, and Runway. Send images and documents as inputs, get validated structured output, and run web research that comes back with citations. Uploaded and generated files are scanned, stored encrypted in your workspace, and fetched through download links that expire in 15 minutes, bound to one workspace and one file. Embeddings, short video, realtime voice, and a sandboxed coding agent are built and bounded in the same workspace but have not yet been proven on a live provider run, and we say so rather than counting them as shipped.

Shared memory for your AI tools

AI Hub keeps a durable, versioned working context per scope and publishes it to AI coding tools over the Model Context Protocol, so a session's memory outlives the session. Your workspace is derived from your sign-in, never something you pick by hand. A write over the size cap is refused rather than silently truncated, and a write that would shrink a large document more than fivefold is held back as a likely wrong scope until you confirm it.

That endpoint also answers a quick provider-preference question by task kind; the full capability-gated routing described above runs in the AI Hub app, not on this endpoint.