Is the recommendation measured performance or just my own configuration read back to me?

Right now it is largely the latter, and the README says so. The bundled catalog has no task-family measurements for the configured models, latency classes are priors, and no paired task-completion outcomes have been collected. Frost surfaces those limits in its results rather than presenting configured rankings as evidence, which is honest, but it means the quality floor you are spending against is a number you supplied.

Does --replay really avoid the API call, or does it just cache the answer?

Replay recomputes the decision from the recorded task text and answers using your current configuration, with no provider call. The catch is that a changed question specification requires fresh analysis, so replaying after you edit the bundled questions-v3.json will not reproduce the original run.

What happens if models.dev is unreachable when I want fresh catalog data?

Nothing breaks in the normal path, because routing never fetches those sources. The separate catalog-build command does the refreshing, shows a diff, publishes atomically and retains the previous catalog. You can also skip it entirely and point catalog_file at your own compatible JSON catalog, or keep durable local changes in catalog.overrides.json.

Will a subscription limit actually change what it picks?

Only for profiles that already clear the quality floor. The optional CodexBar integration reads a neutral capacity file that a separate wrapper generates for you, and can exclude exhausted pools, demote them behind peers in the same tier, or be ignored for a single call. Missing, stale, failed or unverified observations stay unknown rather than being guessed, so an empty file just means capacity is not applied.

How do I install and set up Frost?

Install on macOS or Linux with brew install marcus/tap/frost, copy the starter frost.example.toml and catalog files into ~/.config/frost, set TYPESAFE_API_KEY, then run frost config check before frost route.

What routing policies can I choose?

adequate, quality, fast, and relaxed; the last explicitly lowers the quality floor, and flags override the request object, which overrides configuration.

What are the exit codes?

0 for a recommendation or provisional result, 2 for input or configuration errors, 3 for missing context, no match, or conflicting constraints, and 4 for analyzer failures.

CommunityJev Agents & MCP

Frost

A Go CLI that asks Jev to assess a task once, then picks a model, harness or API locally instead of launching it. Policies weigh cost against a quality floor, and every decision can be recorded and replayed.

See highlights

Open github.com

Stars: 0
Language: Go
License: MIT
Published by: Community
Author: marcus
Last verified: Oct 2026

marcus/frost

User repository on GitHub

View on GitHub

A flexible and configurable CLI model router using TypeSafe Jev.

Forks: 0
Last pushed: Updated Sep 2026

Repo stats from the GitHub API, cached Oct 2026.Homepage

About this project

A model router built on TypeSafe. A Haplab project. Give Frost a prompt, anything from a one-line request to a full markdown doc and it recommends a model, a harness or API, and an effort level where supported. It also gives an explanation and alternatives, in readable text or JSON for your agents and scripts.

Quoted from the project's README. Read the full README.

Tagscommunity jevbest catalog-import awesomejev

Works forJev for Classification Jev for Routing

Used inJev for Software & DevTools

Languages: Go

Highlights

One Jev analysis request on the normal path; the routing decision itself is deterministic Go, so selection does not depend on the model.
Policies are adequate (cheapest that clears the quality floor), quality, fast and relaxed, which explicitly lowers that floor.
Catalog facts and prices come from models.dev and SWE-bench Verified through a separate catalog-build command; routing reads local JSON.
Record a run with --record and recompute it with --replay under a different policy without calling the provider again.

Watch out

The README is unusually blunt that recommendations stay provisional: the bundled catalog has no applicable task-family measurements for the configured models, latency classes are priors, and paired task-completion outcomes have not been collected yet, so the configured rankings are not measured performance. Analysis latency of 104-145 ms mean comes from three small pilot runs and excludes process startup and record-file sync, which the README also says is not an end-to-end guarantee. Cost is extrapolated from list price for analysis only, not measured billing and not the cost of running whatever it recommends. Live routing sends the supplied task text to TypeSafe, and recorded files keep the task text, the answers and the decision, which the README asks you to keep local unless you have reviewed them. Configuration is hand-edited TOML plus JSON catalogs, and the starter profiles are examples that assume access your account may not have. Install is Homebrew on macOS and Linux or a release archive; source builds need Go 1.27 or later. Copy reviewed against the repository README on 2026-10-04, not from running it.

Questions about this entry

Short answers drawn from the linked project or page.

More Jev resources like this

jevyoumean

Shell wrapper that intercepts unknown subcommands, reads the target CLI help output, and asks Jev which documented subcommand matches the intent behind the typo.