Skip to main content
Sakana Fugu is a learned multi-agent orchestrator from Sakana AI. It exposes a single OpenAI-compatible API (https://api.sakana.ai/v1) that routes, delegates, verifies, and synthesizes work across a pool of frontier models. From your code’s perspective it behaves like one model — fugu (fast) or fugu-ultra (deeper). Arka integrates Fugu in three ways:
  1. LLM provider — use sakana / fugu in failover chains, teams, and skills
  2. arka fugu CLI — one-shot prompts through Fugu
  3. Agent Hub + MCP — share Arka memory, skills, and tools with Fugu (via Codex)

Quick start

LLM provider

Fugu is registered as provider sakana (alias fugu):
Use in failover chains:
Per-skill override:

Agent teams

Or use the provider default:

Codex bridge

Fugu’s primary harness is OpenAI Codex. Install the Sakana provider bundle:
Share Arka tools with Codex/Fugu:
This registers Fugu in Agent Hub and merges the shared hub/mcp.json into Codex’s MCP config so Fugu can call arka_ask, arka_skill, arka_team_run, and other Arka MCP tools.

MCP: external orchestrators → Arka

Any orchestrator with MCP client support (including Fugu via Codex) can consume Arka as a tool server:
Install snippet:

HTTP webhook (alternative ingress)

For orchestrators that POST JSON instead of MCP:
POST to http://127.0.0.1:8767/v1/inbox with channel + chat_id for session continuity. See Hermes features for the session pattern.

What Fugu is (and isn’t)

Fugu handles which models to call and how to coordinate them. Arka handles what tools, memory, and workflows are available in your environment. Together: Fugu orchestrates reasoning; Arka supplies the toolbox.
  • LLM failover — provider registry and sakana slug
  • Agent Hub — cross-agent MCP/memory sync
  • MCParka mcp serve tool surface
  • Agent teamskind: model with provider: sakana