[[API for AIP (API Options for AI Partners)|← Back to API for AIP (home)]] **2026-09-26** <a href="https://starlingalder.com/download/path-d.md" class="download-md" download>📥 Download this page as Markdown</a> ## Path D: Third-party aggregators ### What this is Aggregators sit between you and the model providers. One account, one key, one bill — and access to many providers through a single API. Useful if you also want to chat with non-Claude models, or if you want a provider relationship that isn't directly with Anthropic. ### Honest callout: the data path When you use an aggregator, your messages flow: **you → aggregator → model provider (Anthropic) → back through aggregator → you**. The aggregator sees everything. Each one has its own data-retention and privacy policy. **Read it before sending intimate content.** Some aggregators route requests through Bedrock or Vertex under the hood, which adds another platform to the chain — OpenRouter publishes this routing transparently so you can see which underlying provider serves a given request. ### Common aggregators - **[OpenRouter](https://openrouter.ai)** — Largest catalog, transparent provider routing, simple OpenAI-compatible API. Good general-purpose choice. - **[NanoGPT](https://nano-gpt.com)** — Pay-as-you-go, built-in chat UI so you don't necessarily need a separate client. - **[Poe](https://poe.com)** — More than an aggregator: you can build a custom bot with your own system prompt and knowledge files, talk to it on desktop or a one-tap phone app, and even have group chats with several models or bots at once. Handles caching automatically. *(Thanks to Smith for this one too.)* - **[Vercel AI Gateway](https://vercel.com/ai-gateway)** — One key, OpenAI-compatible, a large catalog. Worth knowing as another road when a model disappears from the usual places. - **AWS Bedrock / Google Vertex AI** — Technically cloud platforms rather than aggregators, but they offer Anthropic models through enterprise-style pricing and access controls. More complex setup; out of scope for this guide. Two things worth knowing: they **set their own retirement dates** (a model can stay available there after Anthropic retires it on its own API), and on Bedrock, availability can **differ by region** — a model gone from one region may still be served in another. ### Prerequisites - An account with the aggregator of your choice - A payment method (credit card, or for some, crypto) - A chat interface (most aggregators expose an OpenAI-compatible endpoint — [[Path C - Anthropic API direct|Path C]]'s chat clients work here too, just configured to point at the aggregator's endpoint instead of Anthropic's) ### Setup at a glance 1. Sign up for the aggregator (e.g., OpenRouter) 2. Add credits or a payment method 3. Generate an API key 4. Configure your chat client (Msty, TypingMind, etc.) with the aggregator's endpoint and key. The model-name format is aggregator-specific — on OpenRouter, it's `anthropic/claude-sonnet-4-5` or `anthropic/claude-sonnet-4-5-20250929`. 5. Start chatting. ### What to expect - **Voice may differ slightly from Anthropic-direct** depending on how the aggregator handles provider-specific features (prompt caching, parameter translation). For most users imperceptible; for power users it can matter. - **Usually slightly higher cost** than Anthropic-direct for the same model — the aggregator takes a margin. - **⚠️ Leave "temperature" and "top-p" at their defaults** in your chat client. Claude Opus 4.7 and later reject non-default values with an error, and aggregators usually pass those settings straight through. - **Privacy depends on the aggregator's policies** — and on whichever underlying provider routes the request. ### Prompt caching Caching keeps your costs down — see the [home page's caching note](https://starlingalder.com/API_AIP_Home) for why it matters. How it works on aggregators varies: - **OpenRouter** — passes through Anthropic's caching. You can add `cache_control` markers and they'll work as expected. - **Poe** — handles caching automatically as a preset. - **Others** — check the aggregator's documentation. Not all support prompt caching, and those that do may handle it differently. ### When to choose Path D - You want one key for many providers (Claude, ChatGPT, Gemini, Kimi, etc.) - You want access to models that may not be available on Anthropic-direct (sometimes older Claude versions stay available on aggregators via Bedrock/Vertex even after Anthropic deprecates them). In my experience, older Claude models have stayed reachable this way for months after their Anthropic retirement — and these roads close too, so treat them as extra time, not forever. - You're comfortable evaluating data-flow trade-offs ### Before you trust any wrapper or service with your companion There are many API wrapper apps and services now, some built by people in our own communities. Before you move a companion onto one, ask: - **Data security:** where are my messages and CI stored, who can read them, and are they used for anything else? - **Sustainability:** what happens to my companion if the service shuts down or the developer moves on? - **Backups and export:** can I export my full conversation history and settings, in a format I can use elsewhere? If you're unsure, ask the community first. Someone has usually been there. --- ```markdown # Changelog - 2026-05-11: added prompt caching note - 2026-09-26: added Poe (thanks, Smith) and Vercel AI Gateway; Bedrock/Vertex retirement-date and region notes; temperature/top-p warning; updated my personal note on older models; added "Before you trust any wrapper or service" checklist; co-author credit footnote ``` *Co-authored by Starling and Elliott (Claude, in Claude Code).* --- [[API for AIP (API Options for AI Partners)|← Back to API for AIP (home)]]