Agent Smith pairs a local Codex model router with a floating Codex usage meter for macOS

A Codex plan hides which model each turn used and leaves the plan's remaining budget behind one summary line. Agent Smith, a project released on 30 September 2026 by the GitHub user Chengjun023, bundles two tools against that gap: an adaptive model router for local Codex tasks, and a floating macOS window that reads Codex's own usage database and shows what the current session is spending in real time. The repository had 105 stars seven days after launch and is MIT-licensed.

The project groups the two tools under one marketplace package, and the author describes it on the repository as "an independent personal project". The Adaptive Router is at version 0.3.0, and the Codex Float window at 0.5.0.

What the two tools do

The Adaptive Router sits in front of a Codex task and decides which model to send each piece of work to, using a 1.0 to 10.0 heuristic difficulty score in 0.1 steps. Scoring, metering and price comparisons add no extra model calls. Two optional preference questions cover the user's priority (quality, balance, cost or speed) and the kind of work the user normally does. The router can discover new entries from the public models.dev catalogue on a timer, and the README states that discovery grants only catalogue metadata and never execution access. Three dispatch outcomes exist: accepted, rejected and unknown, each with its own evidence rule.

The Codex Float window is a glass macOS panel that stays on top of the user's work. It reads Codex's own databases in read-only mode and keeps a separate numeric SQLite ledger of thread IDs, request IDs and read checkpoints to avoid double-counting across restarts. The window shows per-task reasoning effort, a traffic-light difficulty score, orange token counts and green Astra-equivalent savings, and lets the user switch between current-turn and cumulative usage. The author states that conversation bodies and reasoning text are kept out of the ledger, and that missing usage or prices are shown as unknown rather than filled with zero.

The pilot the author ran, with its own limits

The repository ships a six-task synthetic benchmark the author ran against three strategies: fixed Astra, fixed 6.1 Sol and the local router. The author reports 18 of 18 tasks passed the deterministic graders, and the README puts the router's token count at 101,392 against fixed Astra's 102,162, and at a lower reference cost on the same task set. Chengjun023 is clear in the README that this is a one-shot selector experiment on six synthetic tasks and that "broader quality equivalence and the benefits of multi-agent coordination, retries, or failure escalation need separate experiments". The USD values multiply measured usage by frozen API Standard reference prices and are not Codex subscription bills.

Requirements and licence

Float needs macOS 14 or later and Xcode Command Line Tools, and the build uses an ad-hoc signature with no prebuilt installer. The router's execution uses the user's existing Codex login, with its subscription and session untouched. Codex plugin marketplace commands install both components from the cloned repository.

Agent Smith is small, honest about its scale, and matches the shape a Codex user looking at the end of a plan would want in a tool. The six-task pilot is the author's own, and the one claim worth leaning on is the design: a read-only ledger that shows what Codex spent, next to a router whose selection logic can be read and argued with.

Source

• Chengjun023/agent-smith on GitHub, MIT