Model and effort suggestions for Codex and Claude Code that follow your task as it grows. It asks TypeSafe's JEV model which model and effort fit the work, then tells you in chat. It never changes a setting.
When a Codex session turns into a multi-tenant architecture review, the reply starts with this:
🔶 Model recommendation
JEV recommends gpt-6-astra / high.
Select it in the model picker if useful. No settings were changed.
Who it is for
People who use Codex or Claude Code and keep one model selected for every kind of task. Start a short summary on a big model and the first check can point you to a lighter one. When the same session grows into an architecture review, it points you to a stronger one.
In Claude Code the callout names Claude models and the /model and /effort commands. You make the switch yourself.
Install
You need macOS or Linux, Python 3.9 or later, and a TypeSafe API key, billed separately from your Codex or Claude plan. Both hosts read the same key: export TYPESAFE_API_KEY where the host runs, or on macOS store it in Keychain:
Then add the plugin to your host:
Codex runs the hook only after you trust it: open /hooks in the Codex CLI and trust the codex-model-advisor@model-picker command. Claude Code needs no such step. Start a new session in either host; its first real prompt runs a check.
Or paste this to Claude Code or Codex and let it do the install:
Use it
Checks run on their own. Three prompts change what happens:
model advisor recheckReassesses your last request now and always shows the result.
mute model advisorSilences alerts for this session.
unmute model advisorTurns alerts back on.
How it works
One stdlib-only Python script, advisor.py, runs as a hook before each prompt in both hosts.
When it checks
Your first real prompt, any prompt that brings a new kind of scope (architecture, integration, migration, security, repeated failures), and every fourth prompt.
When it stays quiet
Replies such as "ok", repeats, prompts within 30 seconds of a check, and non-upgrades after the first check. Each suggestion shows once, with at most one upgrade alert every 15 minutes.
It never blocks
After the check (5-second timeout) the turn runs on the model you selected, so a suggestion applies to your next turn. In Claude Code, two local hooks track your active model.
Status
| Version | 0.3.0, released 2026-09-28. The first release, 0.1.0, shipped on 2026-09-20. Releases |
|---|---|
| Hosts | Codex with plugin hooks, and Claude Code 2.1.257 or later. |
| Platforms | macOS and Linux. No native Windows: the session lock uses fcntl. |
| Tested | Unit tests run in CI on Python 3.9 to 3.14 on Linux, and 3.9 and 3.14 on macOS. The chat callout was checked by hand in the Codex desktop app and CLI, and live checks with real TypeSafe calls passed in Claude Code 2.1.278. |
| Not yet | Suggestions have not been compared against human labels, so they come from a classifier and can be wrong. A benchmark is not published yet. |
No host lets this hook switch models: UserPromptSubmit has no model or effort output. Claude Code does not tell hooks your effort level. All limitations
Questions
What leaves my machine?
api.typesafe.ai under your key: the latest prompt (3,000 characters), the first task excerpt, the last four requests, your model and effort, and the previous suggestion. Session files stay local with mode 0600 in ~/.codex or ~/.claude. It never reads your source files or the full transcript, and there is no telemetry. Privacy detailsCan it switch the model for me?
UserPromptSubmit hook, and neither Codex nor Claude Code lets that hook change the active model or effort. You switch in the Codex model picker, or with /model and /effort in Claude Code.What does it cost?
How do I turn it off?
mute model advisor for the current session, or set "enabled": false in the host's model-advisor.json. To uninstall, disable the plugin and delete the files listed in Configuration.