Mistral Vibe: How to Track Token Usage, Cost, and Output for Mistral's CLI Agent

Mistral shipped a terminal coding agent. Mistral Vibe is an open-source CLI that runs in your shell, reads and writes files, runs commands, and works through multi-step tasks under your supervision — the same habitat as Claude Code and Codex, now with a European model behind it. If you have already wired your workflow around an agent in the terminal, Vibe will feel immediately familiar.

It also has the same blind spot every agentic tool has: once it is running, you have almost no idea what it is costing you. This post covers what Vibe is, why its usage is hard to see, and how to measure tokens, cost, and output for it — on its own and next to the other tools your team already runs.

LobsterOne now supports Mistral Vibe as a first-class source. If you just want the dashboard, it is live at app.lobsterone.ai. The rest of this is the why.

What Is Mistral Vibe?

Vibe is Mistral’s open-source coding agent for the terminal. By default it runs on Devstral, Mistral’s open-weight coding model, but it is model-flexible — you can point it at different backends depending on what you have access to. Because Devstral is open-weight, the same agent can run against Mistral’s hosted API, a model on your own infrastructure, or a local deployment, which matters more than it sounds (more on that below).

In the taxonomy of AI coding tools, Vibe sits in the terminal/CLI agent quadrant alongside Claude Code and Codex — not the inline-autocomplete world of Copilot or the IDE-native world of Cursor. We unpacked why that distinction matters more than feature checklists in the AI coding tool buyer’s mental model, and where the CLI agents land in our tools comparison. The short version: terminal agents are high-autonomy, scriptable, and composable, which makes them powerful and makes their consumption invisible. A single Vibe task can fan out into dozens of model calls before it hands control back.

The most interesting thing about Vibe is not the agent loop, which is well-trodden ground by now. It is the provenance. Mistral is a French company, Devstral is open-weight, and that combination gives European teams a path to an agentic workflow that does not route their code through a US-based API. That is a sovereignty story as much as a tooling one, and we wrote it up separately in data sovereignty for AI coding.

Why Vibe Usage Is Hard to See

Vibe is honest about consumption in a way most tools are not — it ships with --max-tokens and --max-price flags that interrupt a session once it crosses a cumulative budget, counting both prompt and completion tokens. That is genuinely useful as a circuit breaker. It is also a blunt instrument, and it does not solve the visibility problem. Three things make Vibe usage opaque:

Tokens are cumulative and per-session. Vibe keeps a running total inside each session, persisted to your local session history. That is fine while a single session is open. It is useless for the question you actually care about: how much did I — or my team — spend across every session this week? Nobody reconstructs that by hand from session logs.

Cost is an estimate, not a statement. A budget cap computed from token counts and list pricing is a worst-case approximation. It ignores caching, it does not reconcile against what you were actually billed, and it resets every session. Knowing you stayed under €5 on this task tells you nothing about your monthly trend.

Multiple sessions, multiple tools. Real developers do not use one agent. They run Vibe for one task, Claude Code for a refactor, Codex for something else, and switch between them within an hour. Each tool keeps its own counter in its own format in its own directory. The fragmentation is the whole problem — it makes the basic questions (what are we spending, is it making us faster, which workflows benefit) nearly impossible to answer. We have written about that fragmentation tax before; Vibe adds a fourth counter to the pile.

The Metrics That Actually Matter for Vibe

If you are going to measure an agent, measure the things that map to decisions. For Vibe, that is a short list:

Tokens, split by direction. Input and output tokens, tracked over time rather than per session. Vibe — sensibly — does not count cache tokens the way some tools do, which keeps the number closer to the work actually performed. A rising input-token trend usually means context bloat; a rising output trend usually means the agent is doing more generation, for better or worse.

Cost, reconciled and trended. Not a per-session estimate but a daily and monthly figure you can put in front of a finance partner. The absolute number matters less than the slope. A team that doubles its agent spend in a month should know that in week one, not at invoice time.

Lines of code, added and removed. Output volume is a weak proxy for value on its own, but in aggregate and over time it is a useful denominator — cost per thousand lines, output per session, the ratio of removed to added. Vibe’s edits are visible in its session history, which means added/removed lines can be derived from the actual changes rather than guessed.

Sessions and tool calls. How many distinct sessions, how long, and how many tool invocations succeeded versus failed. A high tool-failure rate is a coaching signal — it usually means the agent is fighting your repo’s conventions, not that the model is bad. We made the broader case for treating these as signals, not scores, in what acceptance rate tells you about productivity.

None of these require you to expose your code. They are all metadata about the work, not the work itself — which is the only kind of measurement worth doing.

How to Track Mistral Vibe With LobsterOne

LobsterOne reads Vibe’s usage and turns the four metrics above into a dashboard, automatically, without ever touching your source. Concretely:

1. Connect Vibe. In the LobsterOne Mac app, open Settings, choose Add Tool, and pick Mistral Vibe. It mints a key and starts tracking — no SDK, no code changes, no background daemon hovering over your editor.
2. See the numbers. Tokens (input/output), estimated cost, lines added and removed, session counts and tool-call outcomes — trended by day, week, and month, the same way LobsterOne already shows Claude Code and Codex.
3. Compare across tools. Because Vibe lands in the same provider-agnostic dashboard, you can finally answer the cross-tool questions: which agent is cheapest per line, where your spend is actually going, and whether a shift toward Vibe is saving money or just moving it around.

Crucially, LobsterOne is metadata-only by design. It never sees, stores, or transmits source code, prompts, or model responses — only aggregated usage signals. Raw events are deleted after 90 days; daily summaries are kept for trends. That privacy posture is what lets you measure an agent you would never let a third party read.

Vibe Is a Better Story Next to Your Other Tools, Not Instead of Them

The temptation with any new agent is to frame it as a replacement. That is rarely how it plays out. Most teams that adopt Vibe will run it alongside Claude Code and Codex, picking the tool that fits the task — and, for European teams, picking Vibe specifically when sovereignty or model provenance is the deciding factor.

That is exactly the case where measurement earns its keep. If you are adopting Vibe to move work onto a European model, you want evidence that the migration is happening — that Vibe’s share of your token volume is actually rising, that cost did not balloon in the process, and that output held steady. Without a cross-tool view, “we switched to Mistral” is a vibe (forgive me). With one, it is a number you can show a steering committee.

Vibe is a strong addition to the CLI-agent field, and the European-model angle gives it a reason to exist that is not just “another agent.” But like every agent before it, it is invisible until you make it visible. If you are running it — or about to — connect it to LobsterOne and stop guessing what it costs.