Sakana Fugu — Get Started (docs)

Sakana AI‘s onboarding documentation for Fugu: how to get an API key, call the endpoint, pick a model and effort level, and wire it into the Codex CLI or Claude Code. First-party docs, so T1 for the interface; it states no prices, no benchmarks and no context window.

The opening concession

The page’s first sentence is the interesting one: “Although Sakana Fugu is a multi-agent system, it can be used like a standard LLM through the Sakana API, similar to GPT, Gemini, or Claude.” The product ships an agent system behind a model name — see sakana-fugu for what that does to this spoke’s map.

The interface

Endpoint https://api.sakana.ai/v1, OpenAI-SDK compatible, both Chat Completions and Responses, with Responses recommended for “broader compatibility with modern tooling.”
Keys from the console, shown once. Billing modes and rates live on a separate Pricing page not covered here.
Effort via reasoning.effort: fugu-ultra-v1.1 takes high / xhigh / max; fugu, fugu-ultra-v1.0 and fugu-cyber take high / xhigh (max is accepted and silently maps to xhigh). Any other value is rejected — note there is no low or medium, which is a statement about what the system is for.
Built-in tools including web search; the shipped catalog sets supports_parallel_tool_calls and a freeform apply_patch tool type.

Routing is a key-level setting

fugu routes across all supported providers by default. To exclude one, you enable “Fugu custom model pool” when creating or editing an API key and leave only the providers you want. So provider selection is a property of the credential, not of the request — and the default is “all of them.” This is ai-gateway behaviour sold as a model.

Two operational tells

A two-hour idle timeout. The Codex provider block adds three stream-resilience keys on top of a plain provider definition: stream_idle_timeout_ms = 7200000 (2h, explicitly “don’t drop slow turns at Codex’s ~5-min idle default”), stream_max_retries = 5, request_max_retries = 4. A vendor that has to raise a client’s idle timeout by 24× is telling you its turns run for hours. The docs justify the retries by noting Fugu is stateless, so they’re idempotent.

Shipped agent-conduct instructions. The model catalog’s base_instructions carry safety guards about the agent’s own runtime: don’t run a command that would stop, restart or replace the environment you’re executing in without warning the user and handing them the recovery steps; never force-kill by raw PID (kill -9, Stop-Process -Force, taskkill /F) because the runtime depends on its own child processes — stop the owning task by name instead. That is a provider shipping guardrails inside the model definition, which is the instruction end of the enforcement-vs- instruction axis ../agentic-tooling-wiki tracks on agent-guardrails (and the exact thing nono argues has to be enforced rather than asked for).

Harness integration

One-line installer (curl -fsSL https://sakana.ai/fugu/install | bash) for Ubuntu and macOS, manual setup on Windows, pinning a specific Codex CLI version from the bundle. Claude Code works too, with the docs conceding that because it is closed source there are compatibility differences (“mostly cosmetic”), and warning that the claude-fugu launcher does not auto-update — you re-check this page and apply updates by hand.

What it doesn’t say

No pricing, no benchmark, no context window, no list of which providers are in the routing pool, and no evaluation of any kind. Every capability claim here is a description of an interface, not of performance.

sakana-fugu · sakana-ai · ai-gateway · llm-api-pricing · effort-level · synthesis