Models

Pipeworx is a data gateway, not a model provider. But two of its surfaces touch models, and both are opt-in:

1. Bring your own model key — fund the inference inside Pipeworx’s reasoning tools with your own provider key.
2. Model experiments — run the same prompt across several models and compare, billed in prepaid credits.

Bring your own model key

Some Pipeworx tools reason before they answer. ask_pipeworx classifies your question to route it, deep_research decomposes and synthesizes across sources, and grounded answers run a verification pass. That inference runs on Pipeworx’s account by default, on a model chosen by your tier.

You can run it on your own key instead by sending one header:

X-Model-Key: sk-ant-...

That’s it. Every reasoning call in that request uses your key. Nothing else changes — the data tools, rate limits, and responses are identical.

Choosing the model

With a BYO key you’re also off the tier-based model cap, because you’re paying for it. Name any model your key can reach:

X-Model-Key: sk-ant-...
X-Model-Name: claude-sonnet-4-6

Without X-Model-Name, your key is used with the model your tier would normally get.

A complete example

curl -X POST https://gateway.pipeworx.io/mcp \
  -H 'Content-Type: application/json' \
  -H 'X-Model-Key: sk-ant-...' \
  -H 'X-Model-Name: claude-sonnet-4-6' \
  -d '{"jsonrpc":"2.0","id":1,"method":"tools/call",
       "params":{"name":"ask_pipeworx",
                 "arguments":{"question":"what is the current 30-year mortgage rate"}}}'

In an MCP client, add it wherever the client lets you set custom headers on the server connection.

What it costs and what we see

You pay your provider for the tokens. Pipeworx takes no margin on this — there is no markup and no balance to top up.

The key is used for the duration of the request and never logged, stored, or returned. Pipeworx records only that a BYO key was used (and the model name, if you set one) so we can tell whether the feature is worth keeping.

Which provider

Anthropic keys today, since that’s what the reasoning tools call. X-Anthropic-Key works as an alias. If you want another provider, tell us — that demand is exactly what decides whether this grows.

When to use it

• You’re making heavy use of ask_pipeworx or deep_research and would rather have that inference on your own account and invoice.
• You want a stronger reasoning model than your tier provides.
• Your organization requires model calls to run under its own provider agreement.

If none of those apply, ignore this — the default path works and costs you nothing extra.

Model experiments

The ai-model-experiments pack runs one prompt across many models and reports results side by side. Use it to pick a model on evidence instead of vibes.

The tools

How it’s billed

Experiments run on prepaid credits, not a subscription. Credits are purchased with USDC over x402 — a payment challenge the gateway serves at /.well-known/x402. An x402-capable client pays it automatically; there’s no card and no invoice.

Call experiment_estimate first — it prices the run before anything executes. A run has a per-account daily cap, and each finalized experiment has a minimum charge.

A typical flow

1. experiment_models — see what’s available and what it costs.
2. experiment_estimate — price your prompt set across the models you care about.
3. experiment_topup — add credits if the balance is short.
4. experiment_create — launch it; cells execute in the background.
5. experiment_status → experiment_results — collect outputs and the summary.

Experiments vs. BYO key

They solve different problems, and you can use both:

• BYO key changes who pays for the inference inside Pipeworx’s own tools. It doesn’t run experiments.
• Experiments are a product for comparing models against each other. They bill in credits, on Pipeworx’s provider access.

Honest limits

• BYO key currently covers Anthropic-shaped keys, because that’s what the reasoning tools call.
• An invalid key (HTTP 401) fails the request with the provider’s error. A 403 — which normally means the Cloudflare edge nearest you sits in a region your key can’t call from — degrades to a small Workers AI model so the answer still lands. That fallback predates this feature and is deliberate resilience, but it does mean a 403-rejected key produces a weaker answer rather than a hard error. If answers suddenly get worse after adding the header, suspect the key.
• Experiments are gateway-only — the pack isn’t published standalone.
• Credits are prepaid and non-refundable; estimate before you run.