Switchyard docs

Four steps. There is no SDK to install. You keep the client you already use and point it somewhere new.

1. 01

   Create a workspace and a key

   Sign up, create a workspace, and generate a key. It starts with sy_live_ and is shown once, so put it in your secret manager straight away.

2. 02

   Add provider keys (optional)

   Paste the Anthropic, OpenAI, Google, or OpenRouter keys you already pay for. Skip this step and every request runs on the Yard engine, which needs no provider key.

3. 03

   Rank your tracks

   Pick a model for each track, strongest first: Track 1 Express, Track 2 Mainline, Track 3 Local. Track 4 is always the Yard engine.

4. 04

   Send a request

   Point any OpenAI- or Anthropic-compatible client at Switchyard and send switchyard as the model name. The response headers tell you who answered.

from openai import OpenAIclient = OpenAI(    base_url="https://api.switchyard.space/v1",    api_key="sy_live_...",)resp = client.chat.completions.with_raw_response.create(    model="switchyard",    messages=[{"role": "user", "content": "Explain this stack trace."}],)print(resp.headers["x-switchyard-model"])  # who answeredprint(resp.parse().choices[0].message.content)

curl -i https://api.switchyard.space/v1/chat/completions \  -H "Authorization: Bearer $SWITCHYARD_KEY" \  -H "Content-Type: application/json" \  -d '{    "model": "switchyard",    "messages": [{"role": "user", "content": "Write a commit message for this diff."}]  }'

The -i flag prints response headers, so x-switchyard-model and x-switchyard-track show up above the body.

Every request carries your Switchyard key. Send it in either header below. Both work on every endpoint.

Keys are stored as hashes. We show a key once, when you create it, and can't show it again. If you lose one, create a new one.

To rotate, create a second key, move your clients over, then revoke the first. A revoked key gets 401 on its next request.

Switchyard speaks the OpenAI and Anthropic formats. Connecting a tool means changing two settings: the base URL and the key.

Claude Code

Set two environment variables before you start a session. Claude Code sends its requests to Switchyard, and your tracks decide who answers.

export ANTHROPIC_BASE_URL=https://api.switchyard.spaceexport ANTHROPIC_AUTH_TOKEN=sy_live_...claude

Put both exports in your shell profile to keep them across sessions.

Everything else

Base URL is https://api.switchyard.space. Every endpoint takes your key in Authorization or x-api-key and returns JSON.

Response headers

Every response carries these, streamed or not. Log them and you always know where a request went.

HTTP/2 200content-type: application/jsonx-switchyard-model: claude-sonnet-5.5x-switchyard-track: 2x-switchyard-score: 38x-switchyard-budget-remaining: 6.42

Tracks are your ordered list of models. Track 1 is the strongest model you're willing to pay for. Track 4 is the Yard engine, and it is always there.

How scoring works

Before a request leaves, a small classifier gives it a difficulty score from 0 to 100. It reads the prompt, the attached tools, the context size, and the size of any diff. Scoring adds about 40 ms.

The request then goes to the highest track that still has budget for that score. The score comes back in x-switchyard-score and shows on the departure board.

Budgets move the threshold

Each track has a daily dollar budget. Early in the day, most requests stay on Track 1. As its budget is spent, the score needed to stay there rises. Light work moves down to Mainline and Local, and the hard problems keep the strongest model.

Conversation pinning

A conversation stays on one model. Changing models halfway through a session costs consistency, so Switchyard only switches when the pinned model is out.

What counts as out

• Budget. The track's daily dollar budget is spent.
• Rate limit. The provider returned 429. Switchyard reads the reset headers and uses the track again once the window resets.
• Context. The conversation no longer fits in the model's context window.

When a model is out, the request is switched to the next track down. The departure board logs every request, switched or not.

The Yard engine is gpt-oss-120b running on Switchyard's own GPUs. It is Track 4 on every account and works with zero provider keys, so a new workspace can send requests on day one.

It takes whatever tracks 1–3 can't. When every provider budget and rate limit is spent, your tools keep getting answers.

There is no subscription. Yard engine usage is pay-as-you-go from your credit balance: $0.40 per million input tokens and $1.60 per million output tokens. Credit comes from burning $YARD on Solana from a linked wallet, and the same credit buys the Warm engine, Long context and Wide batches upgrades.

Batch runs

Hand the Yard engine a list of prompts and collect the results later. Start, check, and cancel batches with the MCP tools batch_start, batch_status, and batch_cancel.

Switchyard runs an MCP server at https://api.switchyard.space/mcp. Connect it and your agent can route a prompt, run a batch, or check its credit without leaving the session.

claude mcp add --transport http switchyard https://api.switchyard.space/mcp \  --header "Authorization: Bearer sy_live_..."

Other MCP clients with HTTP transport connect the same way: that URL, plus your key as a Bearer token.

Errors come back in the format of the endpoint you called, OpenAI or Anthropic, with a message that says what went wrong.