Smart Model Router · Codex and Claude CodeDon’t pay flagship prices for every request.
SMR classifies each request into one of four tiers so easy work runs on a light model and only hard work reaches a strong one. Your models and API keys stay on your machine.
- Windows · macOS · Linux
- Open source, Apache-2.0
- Classification from 1 ₫ per request
- Top up from 20,000 ₫, no international card
The problem
The strongest model is stuck doing the small jobs.
A coding agent keeps one model for every request: recolouring a button, writing a test, fixing a typo, and the hardest refactor of the month. You pay for capacity you never use, and the invoice never shows where it went.
-
Small job, big model
A few-line change still goes to the most expensive model because nobody wants to stop and switch models mid-session.
-
Cost with no trace
The model bill lumps everything together. You cannot tell which request should have taken the cheap path.
-
Scattered configuration
Codex and Claude Code use different formats, environment variables, and profiles.
How it works
One routing decision per request.
SMR runs in the background on your machine. It does not call the model for you, does not hold your provider key, and does not rewrite your conversation.
-
1
The agent calls SMR
Codex or Claude Code points at the daemon on 127.0.0.1 instead of calling the model provider directly.
-
2
SMR sends a slice of state
Only a passage of text describing the current request goes to api.acmeai.vn.
-
3
ACME AI returns a tier
The result is one of four tiers: Easy, Lower medium, Upper medium, or Hard.
-
4
SMR calls your model
The original request goes to the model you configured for that tier, using the API key on your machine.
SMR is a local daemon, not an official Codex or Claude Code plugin. It only changes host configuration after you review and confirm.
Four tiers
Four difficulty tiers, four model profiles.
You decide which model handles which tier. SMR never picks a model outside the four profiles you set.
-
Easy
Light model
Small, well-defined change
-
Lower medium
Balanced model
Bug fixes, tests
-
Upper medium
Strong model
Multi-file refactors
-
Hard
Strongest model
Design, complex debugging
Control
The decisions stay on your machine.
-
You choose the model for each tier
Configuration lives in one local file. Changing a model is a config edit, not a reinstall.
-
API keys never leave your machine
Model provider keys stay in local configuration. ACME AI never receives or stores them.
-
Confirm before anything changes
The installer detects your Codex and Claude Code configuration, shows the change, and writes only after you agree. Backups and rollback included.
Cost
Pay for the classifications you actually use.
No subscription, no monthly commitment. Each classification is one line on your ledger. A typical classification costs 1–5 ₫; a cache hit is exactly 1 ₫.
-
1–5 ₫ per request
A typical classification costs 1–5 ₫. An unusually long state can reach about 10 ₫. Amounts round up to whole đồng: no fractions, no hidden fees.
-
Cache hits cost 1 ₫
Re-classifying the same content while the cache is still valid bills 10% of the original price, with a 1 ₫ minimum — for most requests that is exactly 1 ₫.
-
Top up by QR
Bank transfer in VNĐ, no international card. The QR code is shown for 30 minutes; a later transfer is still credited to your balance.
Model costs are still paid to your model provider at their prices. SMR does not replace that bill; it makes sure most requests take a cheaper path. The 1–5 ₫ range and the roughly 10 ₫ ceiling were measured against the production API on 2 November 2026, using Vietnamese state between 500 and 11,800 characters.
Data
Send little, store little, say so plainly.
What SMR sends
A passage of text representing the current request. Not the whole conversation, not tool output, not the system prompt.
That passage goes to the classification service to obtain a difficulty tier.
What is stored
Per request: tier, token count, amount charged, cache status, and latency. Prompts and conversation content are never stored verbatim.
Per account: the email verified by Google or GitHub, display name, and balance.
Claude Code only works with an Anthropic API key or a compatible token; Claude Pro OAuth does not go through SMR. Codex needs a model provider API key; a ChatGPT Plus session will not work.
Questions
What people ask before starting.
Do I need an international card?
No. Top up by VNĐ bank-transfer QR; the smallest package is 20,000 ₫.
Does it work with Claude Pro?
No. Claude Code needs an Anthropic API key or a compatible token. Claude Pro is a signed-in session, not an API key, so it does not go through SMR.
Does it work with ChatGPT Plus?
No. Codex needs a model provider API key. SMR does not use a ChatGPT Plus session.
Does SMR force me to use a particular model?
No. You map models to the four tiers. SMR only reads the difficulty tier and selects the model you assigned to it.
Will installing break my current configuration?
The installer reads your current Codex and Claude Code configuration, shows the changes, and waits for your confirmation. It backs up, can roll back, and lets you skip either host.
What if I transfer after 30 minutes?
The QR code only stops displaying after 30 minutes. A late transfer is still credited to your balance, and any overpayment stays there too.
Install SMR with one command.
The installer places the binary, registers a background service for your account, and opens the local setup page. No admin rights required.
Open the GitHub repocurl -fsSL https://raw.githubusercontent.com/acmeai-vn/SMR/main/install.sh | shirm https://raw.githubusercontent.com/acmeai-vn/SMR/main/install.ps1 | iex