@thinkyou0714/fugu

Zero-dependency TypeScript client + CLI for the Sakana Fugu OpenAI-compatible API.

Fugu is Sakana AI's multi-agent orchestration system exposed as a single model behind one OpenAI-compatible endpoint: you send one request and Fugu decides internally whether to answer directly or to delegate to and synthesize a team of expert frontier models.

Status: unit-tested offline (mocked fetch) + e2e against a local socket stub. A live call against api.sakana.ai needs a real SAKANA_API_KEY — supply one and run npm run smoke to confirm end-to-end.

Highlights

• Zero runtime dependencies. Built-in fetch; ESM-only; Node >= 22.9.
• Typed error hierarchy (FuguAuthError, FuguRateLimitError, FuguTimeoutError, …) with code / status / requestId / retryAfterMs / isRetryable.
• Secret-safe: the raw response body is never stored on errors or logged; messages and --json output are redacted (Bearer … / sk-… / deny-listed keys).
• Typed usage + cost, including Fugu's hidden, billed orchestration tokens.
• Effort-scaled timeouts (reasoning.effort ∈ high | xhigh | max) and status / incomplete / finishReason surfaced (no silent empty strings).
• Resilient transport: retries with exponential backoff + jitter honoring Retry-After, a stable Idempotency-Key across retries, timeout/abort/network classification.
• Streaming (respondStream / chatStream) over SSE — zero-dep.
• Cost controls: a BudgetGuard spend circuit-breaker, output-token + input-size caps, chooseModel() routing (fugu ↔ fugu-ultra), and an opt-in request cache (MemoryCache, LRU+TTL) — identical calls skip the network and the budget charge.
• Strategy patterns: a confidence-gated Cascade (cheap model → judge → escalate), an eval harness (runEval + neutral llmGrader) to measure quality/cost/latency, and concurrency primitives (WorkPool with priority lanes, SingleFlight de-dup) for bulk runs.
• Tool calling (tools + runTools agentic loop, built-in web_search), structured output (respondJson with validate-and-repair), stateful Conversation chaining, and observability hooks (onRequest / onResponse / logger — wire pino/OpenTelemetry; the core stays dependency-free).
• Ecosystem: a multi-provider FuguRouter (failover) and an OpenAI-compatible proxy (createProxyServer / fugu-proxy bin) so Cursor / n8n / any OpenAI-SDK tool can target Fugu — with failover — at a localhost endpoint; a Fugu MCP server (integrations/mcp/, fugu_respond / fugu_chat) for Claude Code / Cursor / Codex; a zero-dep fugu-obsidian CLI (integrations/obsidian/) that answers questions about your Obsidian notes via the Local REST API; an n8n community node (integrations/n8n/) — a declarative Fugu node + Fugu API credential; and a Claude Code /fugu skill + fugu subagent (.claude/, mirrored as a Cursor command) that routes "second opinion" / adversarial-review asks to the MCP server.

Install

npm install @thinkyou0714/fugu        # core (zero deps)
npm install @thinkyou0714/fugu openai # optional: the ./openai adapter

CLI

export SAKANA_API_KEY=...    # from https://console.sakana.ai/get-started
npx fugu "Explain Sakana Fugu in one sentence." --usage
npx fugu "Refactor this..." --model fugu-ultra --effort high
echo "summarize this" | npx fugu --chat
npx fugu --help

--usage prints token usage (incl. orchestration) + estimated cost to stderr; --json prints the (redacted) raw response.

Verify your setup (live)

Once you have a key, confirm a real round-trip works:

export SAKANA_API_KEY=...    # or: cp .env.example .env && edit
npm run smoke                # one minimal call to api.sakana.ai

It prints a PASS/FAIL banner with latency + token usage. On failure it maps the error to a concrete fix (401 → re-copy the key, 403 → plan/model access, 429 → wait, connection/parse → check SAKANA_BASE_URL, timeout → raise/retry). Exit codes: 0 pass, 1 reached the API but failed, 2 not configured (no key) — so CI can skip cleanly.

Troubleshooting

npm run smoke and the CLI map every typed error to a concrete fix. The same diagnosis applies in code — branch on err.code, or use the exported type guards (isAuthError, isRetryable, …):

Error (code / status)	Likely cause	Fix
auth (401)	key missing / invalid / expired	re-copy SAKANA_API_KEY from the console
permission (403)	plan or model not allowed	check tier + model access (fugu-ultra may need a higher plan)
rate_limit (429)	too many requests	wait retryAfterMs, lower concurrency
connection / parse	wrong base URL / non-Fugu endpoint	verify SAKANA_BASE_URL (default https://api.sakana.ai/v1)
timeout	slow run or wrong host	raise the timeout / retry; instant timeouts ⇒ check the host
bad_request (400/404/422)	bad model id or params	check the model (fugu / fugu-ultra) and request body

Programmatic

import { createClient, loadConfig, FuguRateLimitError } from "@thinkyou0714/fugu";

const client = createClient(loadConfig());
try {
  const r = await client.respond("Give me a haiku about pufferfish.", { reasoningEffort: "high" });
  console.log(r.text);
  console.log(r.usage, r.costUsd); // typed usage + estimated cost
} catch (e) {
  if (e instanceof FuguRateLimitError) await sleep(e.retryAfterMs ?? 1000);
}

Streaming + budget + routing

import { createClient, loadConfig, BudgetGuard, chooseModel } from "@thinkyou0714/fugu";

const client = createClient(loadConfig(), { budget: new BudgetGuard({ limitUsd: 5 }) });
const model = chooseModel({ chars: prompt.length, task: "code" }); // fugu ↔ fugu-ultra

for await (const ev of client.respondStream(prompt, { model })) {
  if (ev.type === "delta") process.stdout.write(ev.textDelta ?? "");
  else console.log("\n", ev.result?.costUsd, ev.result?.usage);
}

Tools, structured output, observability

import { createClient, loadConfig, functionTool, webSearchTool } from "@thinkyou0714/fugu";

const client = createClient(loadConfig(), { onResponse: (e) => metrics.record(e) });

// agentic tool loop
const res = await client.runTools([{ role: "user", content: "weather in Tokyo?" }], {
  tools: [functionTool("getWeather", { parameters: { type: "object", properties: { city: { type: "string" } } } }), webSearchTool()],
  handlers: { getWeather: async (args) => fetchWeather((args as { city: string }).city) },
});

// structured output with validate-and-repair
const { data } = await client.respondJson<{ score: number }>("Rate this 1-10", {
  schema: { type: "object", properties: { score: { type: "number" } }, required: ["score"] },
  validate: (v) => {
    const o = v as { score?: unknown };
    if (typeof o.score !== "number") throw new Error("score must be a number");
    return o as { score: number };
  },
});

Cache, confidence cascade & evals

import { createClient, loadConfig, MemoryCache, Cascade, llmJudge, runEval, llmGrader } from "@thinkyou0714/fugu";

// opt-in cache: identical requests are free (no network, no budget charge)
const cache = new MemoryCache({ maxEntries: 1000, ttlMs: 60 * 60_000 });
const client = createClient(loadConfig(), { cache });

// confidence cascade: try fugu, escalate to fugu-ultra only when the judge isn't sure
const cascade = new Cascade(client, {
  stages: [{ model: "fugu" }, { model: "fugu-ultra", effort: "high" }],
  judge: llmJudge(client, { threshold: 0.7 }), // or the zero-cost default `statusJudge`
});
const { result, stageIndex, escalations } = await cascade.run("A hard question…");

// evals: measure quality/cost/latency over a golden set (neutral judge ≠ system under test)
const report = await runEval(client, goldenSet, { grader: llmGrader(judgeClient), concurrency: 4 });
console.log(report.passRate, report.totalCostUsd, report.avgLatencyMs);

Router + proxy (use Fugu from any OpenAI-SDK tool, with failover)

import { FuguClient, FuguRouter, createProxyServer, loadConfig } from "@thinkyou0714/fugu";

const router = new FuguRouter({
  providers: [
    { name: "fugu", client: new FuguClient(loadConfig()), model: "fugu-ultra" },
    { name: "backup", client: new FuguClient({ ...loadConfig(), baseUrl: "https://backup/v1" }) },
  ],
});

// one localhost OpenAI endpoint any tool (Cursor, n8n, scripts) can target:
createProxyServer({ backend: router, token: "local-secret" }).listen(4141);
// or just run the bin:  npx fugu-proxy   ->   http://localhost:4141/v1

Optional OpenAI-SDK adapter

import { createFuguOpenAI } from "@thinkyou0714/fugu/openai"; // requires `openai` installed
const openai = await createFuguOpenAI();
const res = await openai.responses.create({ model: "fugu-ultra", input: "hi" });

Develop

Source runs directly via node --experimental-strip-types — no build needed for dev/tests.

npm test            # offline (mocked fetch) + real timeout wiring
npm run smoke       # live: one real round-trip to api.sakana.ai (needs SAKANA_API_KEY)
npm run coverage    # tests + coverage thresholds
npm run typecheck   # tsc --noEmit (erasableSyntaxOnly)
npm run lint        # biome
npm run build       # tsdown -> dist (ESM + .d.ts)
npm run check:exports  # build + publint + are-the-types-wrong (esm-only)

Layout

fugu/
├── src/
│   ├── index.ts        # public API barrel (the supported entry point)
│   ├── config.ts       # env loading, defaults, effort-scaled timeout
│   ├── errors.ts       # typed FuguError hierarchy + HTTP→error mapping
│   ├── redact.ts       # secret redaction
│   ├── types.ts        # FuguResult / FuguUsage + tolerant parsers
│   ├── pricing.ts      # price table + cost estimation
│   ├── fugu-client.ts  # FuguClient: respond() / chat() / *Stream()
│   ├── retry.ts        # backoff + jitter + Retry-After
│   ├── budget.ts       # BudgetGuard spend circuit-breaker
│   ├── routing.ts      # chooseModel() policy
│   ├── stream.ts       # SSE parsing
│   ├── tools.ts        # tool-calling types + parsing
│   ├── json.ts         # loose JSON extraction (structured output)
│   ├── observe.ts      # logging / metrics hooks
│   ├── conversation.ts # stateful Responses chaining
│   ├── cache.ts        # request cache (MemoryCache LRU+TTL; RequestCache)
│   ├── pool.ts         # WorkPool (bounded concurrency) + SingleFlight
│   ├── cascade.ts      # confidence-gated model cascade + judges
│   ├── evals.ts        # golden-set eval harness + graders
│   ├── router.ts       # multi-provider failover (FuguRouter)
│   ├── proxy.ts        # OpenAI-compatible proxy (bin: fugu-proxy)
│   ├── cli.ts          # CLI (bin: fugu)
│   └── openai.ts       # optional ./openai adapter
├── test/               # client / timeout / p2-p5 / mcp / obsidian / strategy tests
├── integrations/
│   ├── mcp/            # Fugu MCP server (own package; @modelcontextprotocol/sdk + zod)
│   ├── obsidian/       # fugu-obsidian CLI (own package; zero-dep, Local REST API)
│   └── n8n/            # n8n-nodes-fugu (own package; declarative node + credential)
├── .claude/           # /fugu skill + fugu subagent (delegate to the MCP server)
├── .cursor/           # mirrored Cursor command
├── .github/workflows/  # ci / release (changesets + npm OIDC) / codeql (templates)
└── tsdown.config.ts · biome.json · .changeset · tsconfig.json

Publishing

Releases use Changesets + npm Trusted Publishing (OIDC, no long-lived token, automatic provenance). Add a changeset per user-facing change with npm run changeset. See .github/workflows/release.yml.

API reference (as of 2026-06)

Compiled from Sakana's site/console and public guides (see Sources); not verified against a live call in this repo — confirm against your console dashboard.

Item	Value
Base URL	https://api.sakana.ai/v1 (copy the exact value from your console)
Models	fugu (fast), fugu-ultra (max quality)
Auth	Authorization: Bearer $SAKANA_API_KEY
Endpoints	/responses (recommended), /chat/completions, /models
Effort	reasoning.effort ∈ high / xhigh / max

Sources: Sakana Fugu · console · Apidog guide