Nautili
Documentation

Connect Coding Tools

Pointing Claude Code, Codex, Cursor and Continue.dev at the gateway with a base-URL swap, on your existing subscription.

Connect Coding Tools

The coding-tools track points developer surfaces at the gateway with a base-URL swap. Your provider subscriptions keep paying -- the gateway observes and optimizes, and your existing key or seat token is forwarded byte for byte. (The exception is a Nautili issued key: set as the tool's API key, it takes the keyed path instead, with routing profiles, guardrails and Nautili billing. The Claude Code and Cursor tabs below describe both modes.)

This is passthrough traffic. The request is carried byte-identically in both directions on your own credential, with no DevZero header added and no DevZero header returned. It is deliberately outside the published keyed contract.

What you get from it

Passthrough traffic is still telemetry. You get per-request token counts, cost, latency, cache outcomes, session reconstruction and per-developer attribution -- and on a subscription seat the value is denominated in rate-limit headroom rather than dollars, because no bill moved.

What you do not get on a passthrough credential is the keyed product path: routing profiles, guardrail enforcement and the keyed surfaces all require a DevZero issued key. See API access for that.

Base URLs

Substitute your own gateway address. In-cluster that is the gateway service; on Docker or a local binary it is localhost on the port for the surface you want.

Claude Code speaks the native Anthropic wire, so it goes to the Anthropic listener.

On an issued key (two variables). Set the gateway address and a Nautili issued key (sk-dz-), then run claude:

export ANTHROPIC_BASE_URL="<anthropic-base-url>"
export ANTHROPIC_API_KEY="sk-dz-..."

claude

Use the Anthropic base URL the dashboard shows for your gateway (it is already filled into the Claude Code snippet), with no /v1:

GatewayANTHROPIC_BASE_URL
Shared GatewayThe Shared Gateway's one base host, the same host the OpenAI-compatible SDKs use. It routes /v1/messages to the Anthropic listener by path, so there is no anthropic- prefix.
Hosted Proxy Gatewayhttps://anthropic-<your-gateway-host>. The unprefixed host serves the OpenAI-compatible listener.
Self-hosted, hostname routing (the chart default)https://anthropic-<ingress.host>. Without an ingress, the Anthropic listener is port 8081, for example http://localhost:8081.
Self-hosted, ingress.routing=pathsThe single ingress.host, with no prefix.

That is the whole setup. Unlike the passthrough below, an issued key is the keyed, metered path: its requests are governed and billed through Nautili, not paid by a Claude seat. There are no model settings to add: Claude Code's own model names, whether the default, --model opus|sonnet|haiku, or a full name such as claude-sonnet-5-5, work as they are. The gateway maps each one to its catalog id (for example anthropic/claude-opus-5.5), so you do not need ANTHROPIC_MODEL or the ANTHROPIC_DEFAULT_*_MODEL variables, and you do not need to write author/slug ids. Set ANTHROPIC_MODEL only to pin Claude Code to a specific model. A guardrail allowlist written with catalog ids applies to Claude Code unchanged. See model vocabulary for the mapping and its errors.

On your own Anthropic key or a Claude seat. Set only the base URL, to the Anthropic listener (for example http://localhost:8081):

export ANTHROPIC_BASE_URL="http://localhost:8081"

Your ANTHROPIC_API_KEY stays unchanged. A Claude seat token is a subscription credential and keeps the passthrough; the gateway does not reshape it.

Claude Code does not send DevZero's session header. On the native Anthropic surface the gateway falls back to Claude Code's own client session id, which is precisely the client that fallback exists to serve -- so sessions still reconstruct correctly without any configuration.

The install flow lists the clients it has snippets for: Claude Code, Codex CLI, Cursor, Continue.dev, SDK or curl, and a generic backend service. The snippet you copy from the dashboard is pre-filled with your own gateway's address.

Verification

The coding-tools track verifies on any traffic reaching the installation -- not on a specific key. Open a coding session and make one request; the wizard shows it within a few seconds.

Verification is passive. Nothing in the product will originate a model request on your behalf, so the wizard waits for a request you make.

Optimization on this track

This section is about passthrough traffic (your own key or seat). Traffic on an issued key is keyed and metered, and follows the routing and caching rules of the keyed path.

Coding-client traffic can bypass the caches, and that is intentional: a coding conversation is long, mostly unique, and heavily dependent on the provider's own prompt cache, which a response cache cannot help and can hurt.

What does apply is the subscription-regime value story: tokens the gateway does not spend extend how far your prepaid seats stretch. Any dollar figure attached to that is an estimate and is labelled as one. See AI spend.

Do not send keyed traffic to the wrong listener

The native OpenAI listener declares no keyed product surface. A request carrying an sk-dz- key aimed at it takes the subscription passthrough and gets none of the keyed product path -- no guardrail enforcement, no routing profile, no keyed surface. The spend and allowlist gate on the key still runs, so the traffic is not exempt from governance.

Keyed OpenAI-shaped traffic belongs on the OpenAI-compatible listener; keyed Anthropic-shaped traffic belongs on POST /v1/messages on the Anthropic listener.

On this page