Not a model. A harness.

The FinOptix Harness

FinOptix is an orchestration layer that combines skills + MCP tools + steering/routing + agents into one deployable FinOps AI product — built on the open Tinkuy Labs stack (SSTTQ).

agents.finoptix.dev (live)⭐ GitHub

The layers

Each piece is a published, open-source package — the harness deploys them, it doesn't reinvent them.

🧠

Agents

@tinkuy-agent

4 FinOps agents: Analyst, Architecture Auditor, Cost Investigator, Orchestrator. Tool loop, streaming, deterministic grounding.

🔀

Routing

@styrr-llm

Intelligent steering: picks the optimal model per request (complexity L1-L6) with an ordered fallback chain across providers.

⚓

Budget

@sayay-guard

allow/warn/degrade/block per tier. Credit system, budget caps, and model downgrade suggestions — no surprise bills.

🔭

Observability

@qhaway

Every decision traced: model, latency, tokens, cost. Feedback loop (👍/👎) feeds the adaptive classifier.

📚

Context

@tiderag

Knowledge base of architectures and docs (namespace-scoped) via search_knowledge tool.

🧠

Memory

@memory-layer

sqlite-memory-mcp contract: memory/remember, memory/search, memory/recent. Tiered (HOT/WARM/COLD), D1-backed in the cloud.

🛠️

Tools

@byaml-mcp

9+ real AWS FinOps tools: cost, idle resources, tags, architecture generation.

Memory has tiers

Never bulk-load everything. The agent remembers and recalls knowledge across sessions — tiered like S3 storage classes.

HOTmemory/recent~500 tokWhat did we discuss?

WARMcompact_graph~12K tokGeneral context

COLDmemory/search~1-3K tokWhat about payment-service?

ARCHarchive_old0 tokRetention, still searchable

Same contract as sqlite-memory-mcp — your IDE memory and the harness cloud memory share one mental model.

Intelligent routing (L1-L6)

The FinOps complexity classifier picks the right model for every request — 70% of queries hit cheap models.

Level	Example	Model
L1Trivial	What is RI?	finemma-4b
L2Simple	Check tags on this resource	finoptix-7b
L3Medium	Audit this terraform	finoptix-14b
L4Complex	Full architecture review	finoptix-32b
L5Enterprise	Compare 5 accounts + migration	finomotrix-49b
L6Hybrid	L5 with budget constraints	chained inference

How it's deployed

Cloudflare Worker (free tier) — MCP over SSE, D1 + KV storage, scale-to-zero.

User (Kiro / VS Code / Web)
   │  MCP over SSE  /  REST
   ▼
agents.finoptix.dev  ── Cloudflare Worker (Hono)
   │  auth: fp_live_ API key (KV)
   ├── classify() ──→ L1-L6 complexity (domain classifier, <5ms)
   ├── sayay.check() → budget gate (allow/warn/degrade/block)
   ├── styrr.call()  → model + fallback chain (OpenRouter)
   │     └── provider registry: logical family → servable model
   └── qhaway spans → D1 (free tier, scale-to-zero)
        └── feedback loop → adaptive classifier retrain

Free tier storage

✓ D1 (SQLite): spans + feedback — 5M reads/day
✓ KV: API keys — 100K reads/day
✓ No Durable Objects (no billing risk)

Why a harness?

✓ Routing: 70% of queries hit cheap models
✓ Tools: real AWS data, not hallucination
✓ Swap models in the registry — no logic changes

Connect your IDE

Add one MCP server and ask FinOps questions from any agent — Kiro, VS Code, Claude Code, Cursor.

"mcpServers": {
  "finoptix": {
    "url": "https://agents.finoptix.dev/sse",
    "transport": "sse",
    "headers": { "X-FinOptix-Key": "fp_live_..." }
  }
}

← Back to FinOptix