Mistral Vibe (+Free API): This Free AI Coding Agent is ACTUALLY CRAZY!

本站添加: 2026-05-15

[00:00:02][music] >> Hi. Welcome to another video. So, in this one, I want to talk about Mistral Vibe because this is one of those tools that looks very simple at first, but is actually pretty useful if you like terminal-based AI coding.

[00:00:18]Mistral Vibe is basically Mistral's coding agent that runs in your terminal.

[00:00:22]Think of it in the same category as Claude Code, Gemini CLI, Open Code, Qwen Code, and tools like that. You open it inside a project, it reads the project structure, it looks at your Git state, and then you can ask it to explain code, edit files, run commands, write tests, refactor things, and generally work with your code base.

[00:00:41]It has tools for reading files, writing patches, searching the code base, running shell commands, keeping a to-do list, and asking for approval before it does sensitive things.

[00:00:52]It also has {slash} commands, file references with the at symbol, shell commands with the exclamation mark, and different agents or modes. So, this is actually built like a proper coding agent.

[00:01:03]And now, the more interesting part is that Mistral Vibe can use Mistral Medium 3.5 with high reasoning.

[00:01:10]And it is running through the API on the experiment plan. So, this is the setup I want to use for this video. Now, before we install it, let me quickly talk about the pricing because this is where people can get confused. Mistral has the normal Le Chat plans, and then it also has the API side.

[00:01:27]On the Le Chat side, there is a free plan which gives you access to Le Chat for everyday usage, search models, memories, projects, connectors, and things like that with limits. But, the actual all-day Mistral Vibe access is mainly part of the paid plans.

[00:01:42]The Pro plan is $14.99 spent per month excluding taxes.

[00:01:46]That gives you higher usage limits, more thinking, more deep research, more storage, chat and email support, and importantly, access to Mistral Vibe for agentic coding.

[00:01:56]If you are a student, they also show a student option, so check that if it applies to you.

[00:02:01]Then there is the team plan, which is $24.99 spat per user per month, excluding taxes.

[00:02:08]That is more for teams that want shared workspace features, centralized management, domain verification, data export, more storage per user, and that kind of thing.

[00:02:17]And then there is enterprise, which is custom pricing for bigger companies that need custom deployments, security, SSO, audit logs, and all of that enterprise stuff.

[00:02:28]But for this video, we are not going to start with the paid plan.

[00:02:31]We are going to use the Mistral API experiment plan.

[00:02:35]This is the free API plan that is meant for evaluation and prototyping. It has limited rate limits, but they are quite a bit generous.

[00:02:42]So, if you just want to test Vibe, learn how it works, and run some practical coding tasks, it is a really good way to start without paying.

[00:02:50]There is a catch, though, and I actually think Mistral is pretty clear about it.

[00:02:53]API requests made under the experiment plan may be used to train or improve Mistral's models, so the trade-off is basically this. You get free API access for testing, and in exchange your usage may help improve the models.

[00:03:06]To be honest, I think that is a fair trade-off for a free plan, especially because they are not hiding it. But you should know what that means. Do not use the experiment plan with private client code, sensitive company repositories, secrets, or anything you would not want used for model improvement.

[00:03:21]For personal projects, learning, open-source experiments, and toy apps, it is totally fine in my opinion. If that trade-off is not okay for you, then you should either check the privacy settings, use a paid scale plan, or use a local model setup instead. But for this video, I'm fine using the experiment plan because we are just testing and prototyping.

[00:03:39]Now, let's install Mistral Vibe. First, you need Python 3.12 or newer. If you are on Mac or Linux, the easiest way is to run the one-line installer.

[00:03:48]Once it is installed, go into the folder of the project you want to work on. So, CD into your app, your repo, or whatever you want Vibe to see. Then, run Vibe. If you want to run the setup directly, you can also run vibe-setup.

[00:03:59]That brings up the API key setup flow, which is what I'm showing here.

[00:04:02]The setup screen tells you to grab your Mistral API key from Mistral AI Studio and then paste it into the terminal.

[00:04:08]For the free experiment plan route, go to Mistral AI Studio, open the API keys area in the organization or workspace settings, and create a new key.

[00:04:16]>> [snorts] >> If you're using the Lit Chat Pro route, Mistral also points you to the Codastral or Vibe CLI key page. So, basically use the key page that matches your plan. For this video, since we are using the API experiment plan, the organization API key is the one we care about. So, you can give the key a name, choose the access scope, and then click create new key. Once the key is created, copy it right away because like most API keys, you will not be able to see it again later.

[00:04:42]Then, come back to the terminal, paste the key into Vibe, and press enter.

[00:04:46]After that, Vibe stores it locally in your.vibe folder, so you do not have to paste it every single time.

[00:04:52]And that is basically the setup.

[00:04:54]Install Vibe, create the API key, paste it into the setup prompt, and now you're inside the coding agent.

[00:05:00]Now, how do you actually use it? The simplest way is to just run Vibe inside your project and type what you want. For example, you can say, "Explain the structure of this project and tell me where the main entry points are."

[00:05:11]That is a really good first prompt because it lets the model orient itself before making changes. After that, you can ask it to do a specific thing. For example, find the to-do comments and group them by priority, or add tests for the user authentication flow, or refactor this component so the data fetching logic is separate from the UI, or run the test suite and fix the failing test. You can reference files with the @ symbol. So, if you wanted to look at a specific file, you can type @, then start typing the file path, and it should auto-complete. This is really useful because instead of vaguely saying this file, you can directly attach the file context. You can also run shell commands directly by starting with an exclamation mark. So, {exclamation mark} get status or {exclamation mark} npm test or {exclamation mark} ls.

[00:05:56]That bypasses the agent and runs the command in your shell.

[00:05:59]This is useful when you want to quickly check something without asking the model to decide.

[00:06:04]There are also slash commands. Slash help shows you what is available. Slash model lets you change model settings.

[00:06:10]Slash config lets you update configuration. Slash clear clears the session.

[00:06:15]And if you're using the cloud workflow, slash teleport can move a local session into a cloud sandbox so it can keep running.

[00:06:21]For normal usage, I would recommend a very simple workflow. First, ask Vibe to inspect and make a plan without editing.

[00:06:28]Then, once the plan looks good, ask it to implement one small part.

[00:06:32]Then ask it to run the tests or run the app.

[00:06:34]Then review the diff.

[00:06:36]That is the way to use these coding agents without letting them wander too much. Now, you can also use Vibe non-interactively.

[00:06:42]So, instead of opening the chat interface, you can run vibe {dash} {dash} prompt and then write your instruction. For example, vibe {dash} {dash} prompt analyze the codebase and summarize the main modules.

[00:06:53]This is useful for scripts, automation, or quick checks. There are also options like limiting turns or setting a maximum price, which is very useful if you're using an API plan and do not want a task to run forever. Now, let's talk about Mistral Medium 3.5 because this is what makes Vibe more interesting right now.

[00:07:10]Mistral Medium 3.5 is Mistral's newer flagship model for agentic encoding workflows.

[00:07:15]It launched in April 2026 and it is a 128 billion parameter dense model with a 256K context window. It is also open weights under a modified MIT license.

[00:07:26]That is already pretty interesting because a lot of strong coding models are fully closed.

[00:07:31]Mistral says this model combines instruction following, reasoning, and coding in one model.

[00:07:35]It also has configurable reasoning effort, which is why you can see high in the banner here. That means you can use lower reasoning for quick tasks, and higher reasoning for more complex agentic runs where it needs to plan, call tools, inspect files, and make multiple changes. The context window is also important. 256k context is a lot, and for a coding agent, that matters more than people think. It means the model can keep more of the repository, conversation logs, errors, and instructions in view. That does not magically make it perfect, but it does make it much better for multi-file tasks than a tiny context model.

[00:08:09]Mistral also claims strong coding benchmarks for it, including 77.6% on SweBench verified, and they position it as the model that makes remote coding agents in Vibe practical. So, the point is not just that it can answer coding questions. The point is that it is built for long horizon work, tool calling, structured outputs, and agentic workflows. So, for API pricing, Mistral lists Medium 3.5 at $1.50 per million dollar output tokens. So, if you move beyond free experimentation and start using it seriously through API billing, keep an eye on token usage. This is another reason I like starting with the experiment plan first.

[00:08:45]Test the workflow, see if it is actually useful for your kind of coding, and only then think about paying.

[00:08:50]Now, Mistral also published a post called Remote Agents in Vibe powered by Mistral Medium 3.5, and this is actually the bigger story here. Because the idea is not just that Vibe runs locally in your terminal anymore. The idea is that coding agents can now run in the cloud, keep working while you're away, and even run in parallel. So, instead of having one coding agent sitting in your terminal and blocking your whole flow, you can start remote Vibe sessions from the CLI or directly from lit chat. You give it a task, it runs in an isolated cloud sandbox, and you can go do something else while it works. Then, when it is done, it can come back with diffs, progress, questions, or even a pull request.

[00:09:27]This is really important because a lot of coding agent work is not actually about you watching every keystroke.

[00:09:32]If the task is clear, like add tests for this module, fix this CI failure, upgrade this dependency, or refactor this small service, then you mostly want to review the result. You do not want to sit there babysitting the agent for 30 minutes.

[00:09:45]Mistral also says you can teleport a local CLI session to the cloud. So, imagine you start working in the terminal. Vibe already has the context, the conversation history, the task state, and the approvals. Then you realize this might take a while. Instead of keeping your laptop session open, you can move that session to the remote runtime and let it continue there.

[00:10:03]That is pretty cool because it turns the terminal agent into a more async workflow. They are also connecting this to Le Chat. You can start coding sessions from Le Chat, and those sessions use the same remote runtime.

[00:10:14]So, if you're already discussing a feature, a bug, or a repo in Le Chat, you can turn that conversation into an actual coding task.

[00:10:21]That is the direction a lot of these tools are going in now. Chat is not just chat, it becomes a control panel for agents.

[00:10:27]Another thing they mentioned is integrations. Vibe is meant to sit between the tools developers already use, like GitHub for code and pull requests, Linear or Jira for issues, Sentry for incidents, and Slack or Teams for reporting. So, in a team workflow, the idea is that an issue can become a coding session. The coding session can create a branch or PR, and then the result can be reported back where the team works. Now, to be clear, I would not use this for vague tasks, like improve my whole app or make my codebase production ready. That is how you get messy results. But, for contained tasks, this is exactly where remote agents make sense. Test generation, CI investigation, dependency updates, small bug fixes, module refactors, documentation updates, and things like that are perfect candidates.

[00:11:09]The same announcement also talks about work mode in Le Chat. This is not exactly the same as Vibe CLI, but it is part of the same bigger push. Work mode is basically an agentic mode in Le Chat for complex multi-step tasks. It can do research, analysis, cross-tool workflows, use multiple tools, and keep working through a task with approvals for sensitive actions. So, Medium 3.5 is not only powering coding inside Vibe, it is also becoming the backbone for broader productivity agents inside Le Chat. And this is why Medium 3.5 matters more than just the benchmark number.

[00:11:42]Mistral is clearly positioning it as the model for long-running agentic work.

[00:11:46]Local Vibe gives you the terminal workflow, remote Vibe gives you async cloud coding, Le Chat work mode gives you broader multi-tool productivity, and Medium 3.5 is the model tying those together. In practical terms, what is Mistral Medium 3.5 good for inside Vibe?

[00:12:00]So, the first good use case is code base understanding. If you open a new repo and you do not know where anything is, Vibe plus Medium 3.5 is useful for mapping the project. Ask it to explain the architecture, find the main routes, summarize the database layer, or identify where a feature is implemented.

[00:12:17]This is one of the safest and most useful use cases.

[00:12:20]The second use case is test generation.

[00:12:22]Give it a specific module and ask it to add meaningful tests.

[00:12:26]This is where the model can inspect the existing test style, follow the pattern, write the test, run it, and then fix whatever breaks.

[00:12:32]That is a good agent task because the feedback loop is clear.

[00:12:35]The third use case is refactoring.

[00:12:38]Not huge rewrite my entire app type prompts, but focused refactors. For example, extract duplicated validation logic, split a giant file into smaller modules, move API logic into a service, or clean up error handling. Medium 3.5 is useful here because it can reason across multiple files and keep the goal in mind.

[00:12:55]The fourth use case is bug fixing with logs. Paste the error, ask it to inspect the relevant files, and tell it to reproduce the issue if possible. This works well because Vibe can run commands, inspect output, patch files, and rerun the command.

[00:13:07]The fifth use case is documentation. Ask it to generate a readme section, explain setup steps, document environment variables, or write developer notes after it understands the repo. This is not as flashy as making an app, but it is actually very useful.

[00:13:21]The sixth use case is remote or async coding. Mistral is also pushing Vibe remote agents where tasks can run in a cloud sandbox, keep going while you step away, and come back with diffs or even pull requests.

[00:13:32]I would use that for well-defined tasks like adding tests, fixing CI, upgrading a dependency, or handling a small issue from GitHub. I would not use it for vague prompts like make my app better.

[00:13:41]Give it a tight task and it becomes much more useful. Overall, I think Mistral Vibe is worth trying, especially if you already like terminal first AI coding tools.

[00:13:50]It is not perfect, and I still think you should compare it with Gemini CLI, cloud code, code X, open code, and Qwen code.

[00:13:58]But it is a serious option now, and Mistral Medium 3.5 makes it much more compelling than before.

[00:14:04]So, for this video, I would start with the experiment plan. Run a few practical coding tasks, see how it handles the repo, and then decide if it is worth moving to pro for daily usage. That is the sensible way to test it. Overall, it's pretty cool.

[00:14:18]Anyway, let me know your thoughts in the comments.

[00:14:20]>> [music] >> If you like this video, consider donating through the super thanks option or becoming a member by clicking the join button. Also, give [music] this video a thumbs up and subscribe to my channel. I'll see you in the next one.

[00:14:30]Until then, bye.