$0.00

per message — before you even type

Your AI agent is
bleeding money.

And you don't even see it.

Copy the Tokenfesto →

↓ The Token FinOps Manifesto

We Believe

If you build AI agents and you don't know how many tokens each interaction costs, you are the developer who deploys to production without monitoring.
Tokens are the meter, not the goal. Counting them tells you what the bill is — not what the bill bought.
If you enforce least-privilege for your users but not for your tools, you are the security engineer who uses admin for everything locally.
If you right-size your infrastructure but not your context, you are the FinOps practitioner who ignores their own bill.
If you obsess over per-token price while total spend rises, you are measuring the fuel gauge while the tank runs dry.
The discipline you apply to what you build must also apply to how you build.
Token FinOps is not optional. It's the next layer of engineering maturity.
Start measuring. Start governing. Start proving value. Or watch your margins disappear while wondering what happened.

The Hidden Cost of Context

Every message carries invisible weight. This loads before your actual question:

Context entries
25,000 tokens
Steering files
19,500 tokens
System prompt
6,500
Tool schemas
5,000
KB index
3,500
Your message
15

59,500 tokens of context for a 15 token question.
A 4,000:1 ratio. Like reading a novel to answer "what time is it?"

Monthly cost: $400–500

For a single developer. Using credits. Without noticing.

The Principles

Token FinOps isn't a tool. It's a discipline.

Context Is a Resource

Every file loaded into context pays rent on every API call. If it's relevant in <30% of sessions, don't always-load it.

Memory Has Tiers

Hot (last 10 items, 500 tokens) → Warm (summary, 12K) → Cold (searchable archive). Never bulk-load everything.

Budget Guards First

Your agent needs a circuit breaker. Not "alert after $500" — STOP before exceeding $X/day. Non-negotiable.

Route by Complexity

Not every task needs Sonnet. Simple → free model. Complex → expensive. Match cost to value delivered.

Audit Monthly

15 minutes: what's in context? What's dead? What's my memory size? What's my actual spend? Repeat.

Steering Over Skills

Skills everywhere is not free. Every skill manifest pays rent on every call. Use steering for daily, skills for weekly.

Determinism Over Guessing

Validate output against a strict ontology graph, not another LLM. LLM-as-a-judge costs GPU tokens; graph validation costs fractions of a cent.

Value Over Volume

Tokens are the meter, not the goal. Measure cost per call, not per token — and tie spend to outcomes. The discipline that turns a bill into ROI.

Copy the Manifesto

Two audiences: Agent (behavioral rules for your coding agent — executable) or Build (implementation patterns with TinkuyLabs references).

Full Tokenfesto

All 7 rules — install in any coding agent

1,052 tok

Context Is a Resource

Behavioral rule

102 tok

Memory Has Tiers

Behavioral rule

85 tok

Budget Guards First

Behavioral rule

89 tok

Route by Complexity

Behavioral rule

118 tok

Audit Monthly

Behavioral rule

83 tok

Steering Over Skills

Behavioral rule

85 tok

Determinism Over Guessing

Behavioral rule

101 tok

Value Over Volume

Behavioral rule

118 tok

Install in Your Coding Agent

Three install tiers per platform: Steering (compact rules, always-loaded, ~250 tok), Skill (full manifesto, on-demand via keyword, ~950 tok), or MCP (TokenOps server + tool steering, ~500 tok).

Kiro CLI

Steering 242 tok
Skill 1,093 tok
MCP 212 tok

Claude Code

Steering 237 tok
Skill 1,104 tok
MCP 196 tok

Cursor

Steering 236 tok
Skill 1,097 tok
MCP 185 tok

Any Agent

Steering 242 tok
Skill 1,099 tok
MCP 178 tok

Open Source Tools

We don't just talk about it. We build it.

Sayay

Agent cost guardrails. Daily/monthly/session USD limits. Actions: allow, warn, degrade, block.

npm i @carloscortezcloud/sayay-guard

Styrr

Multi-model LLM router. Automatic fallback chain. Route by cost, speed, or quality.

npm i @carloscortezcloud/styrr-llm

Tinkuy

Minimal agent framework. Tool loops + budget control + observability. 200 lines core.

npm i @carloscortezcloud/tinkuy-agent

SOFE

FinOps governance engine. 36 policies, 18 collectors. Cloud + AI cost evaluation.

pip install sofe

Join the Movement

Token FinOps is year 1. Cloud FinOps took 7 years to become mainstream.
We're starting now.

Read the Blog Series → GitHub →

Built by engineers who got tired of burning tokens.

⌁ Aligned with the Tokenomics Foundation

Tokens are the meter, not the goal.

The Linux Foundation launched the Tokenomics Foundation (Aug 2026) to standardize the economics and ROI of AI. Tokenfesto is its practitioner side — the discipline you apply before the standards arrive.

Consumption → cost-to-serve · Big-T → classify complexity · Value → outcome-linked spend
Tokenomics Foundation ↗

"Tokenomics isn't about counting tokens. It connects the total cost of AI to its outputs." — Tokenomics Foundation