Impulse releases its MCP for Codex
We just released the Impulse MCP for Codex. Connect it once, and your coding agent can work directly with our autonomous ML engineer: send it data and an objective, watch training progress, compare approaches, deploy the winner, and integrate it into your codebase. You never leave Codex.
Install it with one line:
You need an Impulse account at impulselabs.ai, and that's it. Docs are at docs.impulselabs.ai. Claude Code support is coming.
Why we built this
If you build software today, there's a good chance you spend your day inside Codex. Your coding agent has the full context of your project: the codebase, the data, what you're actually trying to ship.
But the moment you need a predictive model, you're stuck. That work normally goes to an ML team, and most teams don't have one. Talented MLEs are expensive and hard to hire, so the predictive feature waits for headcount or dies in the backlog.
That's what this release fixes. Connect the Impulse MCP and your coding agent gets an ML engineer on the team. Impulse does the specialist work, inside Codex, next to everything else you've connected.
We built our Fed model this way
This isn't a demo we cooked up for launch. Our Fed rate prediction model, the one that's been live at fedratewatch.impulselabs.ai since June, was built through the Impulse MCP inside Codex. We gave Codex the objective in the editor, it worked with Impulse over the MCP, and the model came out the other end. We've been using this workflow ourselves for a while. It works.
What Codex can do with it
Impulse is an autonomous MLE agent. You give it data and a problem, and it profiles the dataset, designs a validation strategy, trains and compares multiple approaches, and hands back models, metrics, predictions, notebooks, and training code. It also deploys the winning model and serves inference, so you end up with a working endpoint you can call.
Through the MCP, Codex can:
Send Impulse a dataset and an objective straight from your project
Monitor a long-running training session and check in when something meaningful happens
Pull the full account of a run: which approaches were tried, how each scored, which one won
Deploy the winning model and check deployment status
Prepare prediction rows that match the model's feature contract and run inference
Retrieve every artifact and put it where it belongs in your repo
Why two agents beat one
Why would a coding agent talk to an ML agent instead of doing the work itself?
Because they know different things. Codex has read your codebase and knows your problem. Impulse knows machine learning. So Codex can brief Impulse the way an engineer briefs a specialist: the objective, what the application needs, which data actually exists at prediction time, what to optimize for. You don't have to know that your problem is a multiclass classification task that should optimize log loss. Codex translates your product problem into an ML problem, and Impulse takes it from there.
The same goes for results. When Impulse reports back, Codex reads the metrics against what your product needs and directs the next round: train on this slice, optimize for this instead, deploy that one.
You set the autonomy
You can review every result yourself and tell Codex what Impulse should try next. Or you can tell Codex to keep going until it has a defensible answer and only pull you in when there's a real decision to make. Both work. You decide how much runs without you.
The model ends up in your codebase
This is the part we care about most. Because Codex has your repository, it takes what Impulse built and makes it part of the product: model-loading and inference code, preprocessing that matches the feature contract, API endpoints, tests, config, documentation. And since Impulse deploys and serves the model, done means a live endpoint wired into your application.
Impulse builds the model. Codex ships it.
Getting started
Sign up at impulselabs.ai if you don't have an account, connect the MCP, then give Codex an ML objective and point it at your data. Full documentation at docs.impulselabs.ai.
If you'd rather start outside Codex, the platform works the same way it always has: upload data, write a prompt, get a model, at impulselabs.ai.