If you run more than one AI coding agent at a time — Claude Code in one terminal, OpenCode in another, maybe a side pane open to a model you are evaluating — you already know what context switching costs you. Macchiato's Day 2 release leans into that pain. It is small, iterative, and the direction it picks matters more than any single feature it ships.
What Day 2 Actually Adds
The Day 2 update bundles three changes worth naming:
- A live token and cost sidebar that updates as your agent runs, instead of forcing you to wait for the end-of-session summary.
- A consumption dashboard that surfaces cumulative spend per agent and per session.
- Keyboard shortcuts to switch between Claude Code and OpenCode panes inside a single terminal window.
None of those are headline-grabbing in isolation. Together, they signal what Macchiato is trying to become: a multiplexer purpose-built for agentic CLIs, not a generic tmux clone with a coat of paint.
The token meter is the change you will feel first. Most coding agents only report cost when a session ends, which means an Opus-heavy refactor can quietly burn through a multi-dollar budget before you notice. A live counter shifts that feedback loop from minutes to seconds — closer to a heart-rate monitor than a monthly bill.
Why Token Metrics Belong in the Terminal
There is a real argument for putting cost telemetry one keystroke away from the agent spending it.
When you run a single agent occasionally, a monthly bill is enough oversight. When you run two or three in parallel — Claude Code working a backend refactor while OpenCode triages a frontend bug — that bill stops being interpretable after the fact. A sidebar that splits spend by agent and session gives you a much faster intuition for which workflows are expensive and which are not.
Live token meters are a soft form of guardrail, not a hard one. They surface cost while you still have time to stop a runaway loop, but they will not cap spend or pause the agent. If you need a hard limit, set it at the API key or organization billing level — not in the terminal UI.
This matters more than it might seem. Many teams adopt AI coding agents on a trial budget and then have no clear story for who owns the spend. A per-agent dashboard makes that conversation tractable: you can point at "the Claude Code session that touched the auth module cost $4.20" instead of "AI tools were $312 last month." The first sentence leads to a decision. The second one leads to a meeting.
It also changes how you write prompts. When the meter ticks up in real time, you naturally start trimming context, scoping tasks tighter, and reaching for cheaper models on work that does not need a flagship. That self-correction loop is hard to build any other way. You can read all the cost-control advice you want; nothing teaches faster than watching $0.40 appear on screen during what you thought was a small change.
The Multiplexer Pattern, Quietly
Macchiato's bigger bet is that the multiplexer is the right unit of UX for agentic work. The shortcut-driven switching between Claude Code and OpenCode hints at where this is heading.
You can already run multiple agents in tmux, of course. But generic terminal multiplexers do not know that one pane is Claude Code and another is OpenCode. They cannot aggregate token usage across panes, cannot show per-agent cost, cannot give you a single view of "what are all my agents doing right now." Macchiato can, because it understands its tenants.
The closest analogy is a process supervisor for AI agents: each pane is a workload with its own resource budget, and the wrapper handles observability. The Day 2 release reads like the first step toward that model — get the meter in place, get the switching feel right, then layer on policy and budgets in subsequent releases.
If you are already using a generic multiplexer, you do not need to switch to evaluate Macchiato. Run it in one tmux pane against a throwaway repo for a week and watch how often you reach for the token sidebar. If you reach for it more than once a day, the cost story alone may justify the migration.
It is too early to call any tool the winner in this space. There will be other entrants — terminal-native dashboards, IDE-integrated meters, vendor-specific cost tooling from the model providers themselves. What is clearer is that "I run several agents and I need to see what they are doing" is now a real workflow, and the tools that serve it directly will pull ahead of the ones that treat agents like ordinary CLI processes.
Who Should Care Right Now
Three groups will get the most value from Day 2 today:
- Solo developers running two or more agents in parallel. The token sidebar pays for itself the first time you catch a runaway loop early.
- Small teams piloting AI coding tools. Per-agent cost dashboards make budget conversations concrete instead of hand-wavy, which matters when you are arguing for a larger AI tooling line item next quarter.
- Engineers already living in tmux, zellij, or screen. Macchiato's switching model will feel familiar enough to evaluate in an afternoon. The keyboard shortcuts are the lowest-friction part of the release.
If you only use a single agent on a single project, you can safely wait. The cost meter is useful but not transformative at that scale, and the parallel-terminal story does not apply yet. Day 2 is aimed at people whose workflow has already grown past one agent — not a pitch for adopting more.
What we would watch for next: a hard-cap option on the cost meter (today it is read-only), session-level budget alerts, and support for additional agentic CLIs as the field expands. If Macchiato delivers those without inflating the surface area too much, the Day 2 changes will look in hindsight like the foundation rather than the highlight.
Originally published at pickuma.com. Subscribe to the RSS or follow @pickuma.bsky.social for new reviews.
Top comments (0)