How to Get Your Dev Tool Picked by AI Coding Agents

TL;DR: Developers used to find tools by browsing, clicking, and landing on your page. Increasingly they may not open a browser tab at all: they stay in their AI coding agent, and it picks the library or package for them. The landing-page click still matters, but it's no longer the whole game. There's a new asset emerging alongside it: the agent skill, a standing presence inside the developer's workflow that behaves more like an app install than a pageview. Here's how your tool becomes the one the agent reaches for, and how to start now.

Where developers actually decide now

The developer isn't necessarily out browsing for your tool. More and more of the building happens inside an AI coding agent. In Supabase's 2026 State of Startups survey, 61% of startups now have more than half their codebase written by AI, and Claude Code was the tool they named most. The places you used to catch them are quieter, too: asked which industry events they had attended or planned to attend, 60% picked none of them, and X, LinkedIn, and Reddit each lost a slice of that audience. Picture a developer describing what they want to an agent and letting it assemble the stack. That's increasingly the moment your tool gets chosen or skipped.

The developer stack is going agent-native

The landing page still has a job. Someone comparing vendors will click, read, and sign up. But a growing share of "which tool should I use here" moments now happen inside a coding agent, in seconds, without a browser tab ever opening.

The agent is choosing your tools for you

When a developer tells an agent "add auth" or "set up a database," the agent picks the library. And it picks the one it already knows. Researchers who tested what coding models reach for found they default to the popular option: a familiar library like NumPy even when the task doesn't need it (up to 45% of the time), or Python for 58% of high-performance project setups where it isn't the right language. Those choices ship, too. With most of the code in these startups now AI-written, the tool that gets written in is increasingly the one the model already knows, not the one a developer sat down and researched.

Models will even write in things that don't exist. In an analysis of 2.23 million AI-generated code samples, roughly 440,000 referenced software packages that aren't real. And it's all happening against a real trust gap. In Google's 2025 DORA research, 90% of technology professionals now use AI at work, while 30% report little to no trust in the code it generates. Teams are leaning on AI heavily while holding its output at arm's length. That gap is the reason defaults are worth attention rather than a reason to dismiss them: whatever the agent reaches for is what a developer is left to keep or reject, and you can't be kept if you were never proposed.

The lesson for a dev-tool brand is the same either way: the agent defaults to what it knows. If your product isn't part of what it reaches for, you aren't in the consideration set, and there was never a click to lose in the first place.

Agent skills: a new asset worth owning

This is where agent skills come in. Anthropic defines an Agent Skill as an organized folder of instructions, scripts, and resources that an agent can load to perform a specific task, and in December 2025 it published the format as an open standard, so it isn't Claude-only.

In practice, a skill teaches the agent how to use your product correctly. MongoDB already ships an official Claude Code plugin that bundles its MCP server and prebuilt skills for common MongoDB tasks, so an agent can reach for MongoDB directly while a developer works. Once a developer installs it, it doesn't fire once and disappear. It stays in the workflow, available every time they build.

That permanence is the point. A skill is less like an ad impression and more like earning a spot in the toolbox.

There's a parallel move in documentation: the llms.txt convention, a proposal for a machine-readable version of your docs, aimed at the same problem of an agent working from whatever it can find about your product.

How to get picked

You can't force an agent to choose you. But you can make your tool the obvious thing to reach for. Three moves matter most:

• Make your product usable inside the agent. A skill teaches the agent how to work with your product; an MCP server, once a developer configures it, lets the agent actually operate against it. MongoDB ships both together for that reason. That's the difference between a link the model might mention and something it can use while a developer builds.
• Publish AI-readable docs. A machine-readable version of your docs, along the lines of llms.txt, gives an agent something accurate to work from instead of guessing at your API. Nothing consumes it automatically yet, so treat it as removing an excuse for getting your product wrong rather than as a distribution channel.
• Earn your way into the default set. Models lean on the popular and the familiar, and there's no trick that skips that. Real adoption, clean public docs, and working examples can help put you in the pool the model draws from. The first two moves feed it: the more places your tool works cleanly, the more chances the agent has to reach for you.

None of this guarantees the pick. It stacks the odds, and it puts you in the agent's path instead of hoping a developer finds your landing page later.

Think installs, not pageviews

If you're used to measuring developer ads by clicks and landing-page conversions, this one will feel unfamiliar. The old scoreboard doesn't apply.

• There is no click-tracking pixel on a skill install the way there is on a display ad.
• A skill captures the open-ended "what should I use here" moment, not the developer who already decided and searched your name.
• The ecosystem is early, which is exactly why it's worth moving on now. 57% of the builders Supabase surveyed are already using or trialing MCP, and 52% are building or planning to build agents, so the audience is here even while the ad formats are still forming.

Measure it the way you'd measure a distribution channel: installs, activations, and whether your tool turns up in more of your users' projects over time, not a single last click.

What this means for your Carbon Ads campaigns

None of this means tearing up what already works. The contextual sponsorships you run across developer sites still reach people in a research mindset, and they still convert. Keep them.

What changes is where some of that spend can point. Instead of every placement driving a click that might convert once, a placement can drive a skill install: a standing presence in a developer's daily workflow. Same audience, a different and more durable kind of shelf space.

Carbon Ads has spent since 2010 putting developer brands in front of developers respectfully, on the surfaces where they actually spend their attention. That surface is moving into the agent. If you want to talk through where your placements should point, how to frame an install instead of a click, and how to measure it, reach out to our team. You don't need a full strategy overhaul to start.

Frequently asked questions

What is an agent skill?

An agent skill is a packaged set of instructions and resources that teaches an AI coding agent how to perform a task, such as using a specific product. Anthropic introduced the format for Claude and published it as an open standard in December 2025, so it isn't limited to one vendor's agent. For a dev-tool company, a skill is a way to be present and usable inside the agent, not just discoverable on the web.

How do I advertise inside AI coding agents today?

The formats are still early, so it's a mix. Keep your contextual sponsorships reaching developers in research mode, and start pointing some placements toward a skill or MCP install rather than only a landing page. The goal is to be the tool the agent already knows how to use. Talk to Carbon about where your placements should point and how to measure installs instead of clicks.

Does the landing page still matter?

Yes. Developers comparing vendors still click through, read, and sign up, and that funnel still converts. The shift is that it's no longer the whole funnel: a growing share of tool decisions now happen inside an agent, so the landing-page click is one surface rather than the only one.

A shorter version of this argument first appeared on the BuySellAds blog.