Model gateway (LiteLLM)

SWEny runs on headless Claude Code. It does not ship a native Gemini or OpenAI provider, and it is not going to. But Claude Code respects ANTHROPIC_BASE_URL, so you can point the backend at any Anthropic-compatible gateway (a LiteLLM proxy, a corporate egress proxy, an observability proxy) and let the gateway route to whatever model you want. Combined with per-node model selection, this gives you cost tiering: a cheap model on the grunt steps, a strong model on the reasoning steps.

Tool use and structured output are unsupported through a gateway

SWEny relies on Claude Code’s tool-use protocol and JSON output behavior, which assume an Anthropic model. Routing a node to a non-Anthropic model (e.g. Gemini via LiteLLM) may produce flaky tool calls or output that does not honor a node’s output schema. This path is at your own risk. Use cheaper Anthropic models (e.g. claude-haiku-4-5) for the most reliable cost tiering, and reserve non-Anthropic models for nodes that do not depend on tools or structured output.

Environment variables

Variable	Header	Purpose
ANTHROPIC_BASE_URL	n/a	Base URL of the gateway, e.g. https://litellm.internal/v1.
ANTHROPIC_API_KEY	x-api-key	API-key auth. Real Anthropic console keys are billing-sensitive.
ANTHROPIC_AUTH_TOKEN	Authorization: Bearer	Bearer auth. Most LiteLLM deployments expect this (e.g. sk-litellm-...). Not a console key.
SWENY_AUTH	n/a	Auth mode: auto (default), api-key, or oauth.

Why SWENY_AUTH exists

By default (auto), when a Claude Code OAuth token is present, SWEny strips ANTHROPIC_API_KEY from the subprocess environment. This protects you from a stray .env key silently billing a metered API account when you meant to use your subscription.

That protection gets in the way when you deliberately want to authenticate a gateway with a key while an OAuth token also happens to be in your environment (common after claude setup-token). Set SWENY_AUTH=api-key to opt in: SWEny then preserves both ANTHROPIC_API_KEY and ANTHROPIC_AUTH_TOKEN.

SWEny does not infer this from the presence of ANTHROPIC_BASE_URL. A base URL does not mean “non-billing gateway”: pass-through proxies set it while still billing your real key. The auth choice must be explicit, because guessing wrong toward a key fails silently and costs money, while guessing toward OAuth just fails the call and you fix the config.

SWENY_AUTH	Behavior
auto (default)	When an OAuth token is present, strip ANTHROPIC_API_KEY. Never touches ANTHROPIC_AUTH_TOKEN.
api-key	Preserve ANTHROPIC_API_KEY and ANTHROPIC_AUTH_TOKEN even when an OAuth token is present.
oauth	Force subscription/OAuth: strip both ANTHROPIC_API_KEY and ANTHROPIC_AUTH_TOKEN.

GitHub Action

The action exposes matching inputs:

- uses: swenyai/sweny@v5
  with:
    workflow: .sweny/workflows/upgrade.yml
    anthropic-base-url: https://litellm.internal/v1
    anthropic-auth-token: ${{ secrets.LITELLM_TOKEN }}
    sweny-auth: api-key

Cost tiering example

Put a cheap model on the mechanical nodes and the default (or a stronger) model on the reasoning node. Resolution is node.model → workflow.model → executor default.

id: helm-upgrade
name: Helm chart upgrade
entry: plan
model: claude-opus-4-6 # workflow default for the thinking
nodes:
  plan:
    name: Plan the upgrade
    instruction: Read the chart and the changelog, decide what changes are needed.
    skills: [github]
  apply:
    name: Apply mechanical edits
    instruction: Make the version bumps and formatting changes.
    skills: [github]
    model: claude-haiku-4-5 # cheap grunt work
edges:
  - from: plan
    to: apply

With ANTHROPIC_BASE_URL pointed at LiteLLM, you can alias claude-haiku-4-5 (or any model name) to a cheaper backend in your LiteLLM config.

Verifying the connection

sweny check is gateway-aware. When ANTHROPIC_BASE_URL is set it probes the gateway (not real Anthropic) and reports the selected auth mode. The base URL is redacted to scheme and host in output so a credential embedded in the URL never logs.

✓  Anthropic (gateway)
   gateway reachable (https://litellm.internal), auth mode: auth-token

Caveat: prompt caching

Mixing models across nodes means a cheap grunt node does not share prompt cache with the expensive reasoning node, and a gateway may disable Anthropic prompt caching entirely. The per-node savings are real, but do not assume cache hit rates from a single-model run carry over.