diet-ai — Portion control for your AI agents

How it works

The agent is the router.
Not a regex. Not a proxy.

diet-ai gives your agent a cost policy. The agent — already an LLM — semantically understands each task and picks the cheapest model that fits. No keyword matching. No middleware.

Think of it like a real diet.

You read the diet plan. You understand the rules. But you decide what to eat at each meal — not a calorie-counting app making every choice for you.

Same here. The agent reads .diet.yaml, understands the tiers, and makes its own call: “This is a file search — Haiku is fine” or “This is a security review — I need Opus.”

Scan

Detects your tools — AGENTS.md, CLAUDE.md, .cursorrules and more.

Set the diet

Writes a .diet.yaml policy — fast / standard / premium tiers with cost guidance and example patterns.

Teach the agent

Injects the policy into your agent’s config. The agent reads it, understands the rules, and picks the right model when spawning sub-tasks.

Track

Every routing decision is logged. Run diet stats to see how much you’ve saved.

Why not all tools? The agent needs the ability to delegate tasks to a different model. Copilot CLI’s task(model: "haiku") and Codex support this today. Claude Code and Cursor run a single model per session — no sub-agent routing yet. We’ll add support as soon as they ship it.

.diet.yaml

# 🥗 Generated by diet-ai
profile: coding
default_tier: standard

tiers:
  fast:     # $1.25 / 1M tokens
    models: [claude-haiku-4.5]
  standard: # $9.00 / 1M tokens
    models: [claude-sonnet-4.6]
  premium:  # $45.00 / 1M tokens
    models: [claude-opus-4.6]

rules:
  - name: simple-tasks
    match:
      keywords: [draft, email, format,
                 recap, summary, list]
      agent_types: [explore, task]
    tier: fast

  - name: code-changes
    match:
      keywords: [implement, refactor,
                 fix, debug, build]
    tier: standard

  - name: deep-analysis
    match:
      keywords: [research, architecture,
                 security, strategy]
    tier: premium

Task	Needs	Without diet-ai	With diet-ai
Draft an email	🟩 fast	$45 / 1M tokens	$1.25 / 1M tokens
Format a recap	🟩 fast	$45 / 1M tokens	$1.25 / 1M tokens
Refactor a function	🟡 standard	$45 / 1M tokens	$9.00 / 1M tokens
Review a PR	🟡 standard	$45 / 1M tokens	$9.00 / 1M tokens
Security architecture	🔴 premium	$45 / 1M tokens	$45 / 1M tokens
100 calls/day (avg 2K tokens) — Monthly cost		~$270 / month	~$54 / month

Put your AI agents on a diet

Automatic routing where it works.
Coming soon everywhere else.

The agent is the router.
Not a regex. Not a proxy.

Scan

Set the diet

Teach the agent

Track

Your AI models are eating
your budget alive

Your agents are hungry.
Time to put them on a diet.

Put your AI agents on a diet

Automatic routing where it works.Coming soon everywhere else.

The agent is the router.Not a regex. Not a proxy.

Scan

Set the diet

Teach the agent

Track

Your AI models are eatingyour budget alive

Your agents are hungry.Time to put them on a diet.

Automatic routing where it works.
Coming soon everywhere else.

The agent is the router.
Not a regex. Not a proxy.

Your AI models are eating
your budget alive

Your agents are hungry.
Time to put them on a diet.