JetBrains bets against itself, and that's the winning move

JetBrains just made its biggest bet against itself. Rider 2026.2 now supports third-party providers for AI completion: OpenAI-compatible endpoints, LM Studio, llama.cpp, and even Mercury by Inception Labs. You can run inference locally with certain models, keeping data on your machine.

Here is the strange part. JetBrains sells its own AI Assistant and its own agent, Junie. It also serves roughly 15 million professional engineers, including 88 of the Fortune 100, according to OpenAI's November 2025 profile. So why would a company hand its users the keys to walk away from its own product?

Because the smart money in developer tooling is no longer betting on lock-in. It is betting on openness. And I think that shift changes the rules for everyone building tools right now.

The Orchestrator Advantage

Here is the framework: in the AI IDE war, the winner is not the best model. It is the best orchestrator.

An orchestrator does not force you into one stack. It lets you swap models based on cost, speed, privacy, or task. Think of it like a kitchen. A closed IDE hands you one chef and one menu. An orchestrator hands you the whole kitchen and lets you cook.

The data backs this up. JetBrains' 2026 AI Pulse survey found 90% of developers regularly use an AI tool at work. A separate roundup found 69% of companies run three or more models in production, and 67% of developers use multiple AI tools. Nobody lives in one model anymore.

So the competitive question flips. It is not "whose model is best." It is "who makes switching between models feel like nothing." That is the Orchestrator Advantage. The nicher the workflow you serve, the stickier the shell becomes, even when the models underneath keep changing.

How Rider Actually Splits Agent From Completion

Here is the clever architecture move, and it is worth slowing down on.

Rider 2026.2 separates two things most people lump together: the agent and the completion model. GitHub Copilot ships as a "natively integrated agent," per the July 22, 2026 release post. AI completion, on the other hand, can run through third-party providers you wire in yourself. Two different jobs, two different dials.

Think of the agent as a 500 IQ intern that plans, executes, and iterates across your codebase. Completion is the fast autocomplete whispering the next line. Different tasks, different budgets. Why pay frontier prices for both?

That split is the whole ballgame. It means you can pair a heavy agent for big refactors with a cheap local model for grunt completion. A little setup here saves you a lot of token spend later.

Now the local part. Through LM Studio and llama.cpp, Rider can hit an OpenAI-compatible endpoint on your own machine. Zero data egress. That is the phrase compliance teams have wanted for two years. Your source code, your environment variables, your secrets never leave the building.

Then there is Mercury by Inception Labs, a diffusion-based language model that was launched in February 2025. Most coding models generate text left to right, token by token. Diffusion models sketch the whole answer and refine it, which can be much faster. Seeing a new model class land inside a mainstream IDE this quickly tells you the plug is real, not marketing.

Here is my honest take. This is not pure openness. JetBrains still routes a lot of usage through its own AI service and credit system, and LeadDev reported the AI Ultimate tier at $30 per month with 35 credits per 30 days. The company sits in the middle of the billing. So the lock-in did not vanish. It moved up a floor, into the workflow and subscription layer, where JetBrains still owns the experience.

But moving the lock-in up is exactly the point. Sell the kitchen, not the one dish. The IDE becomes the thing you keep. The models become swappable parts. Simple always defeats complex, and "keep your shell, change your engine" is a very simple pitch.

One caution I will not skip. An open model market widens the attack surface. WithSecure documented a prompt injection flaw in Rider's AI Assistant, and CERT-EU flagged CVE-2024-37051 at CVSS 9.3, where JetBrains IDEs could leak GitHub tokens. More endpoints means more places for a bug to hurt. Openness is a strategy, not a free lunch.

Three signals inside the same shift

2031

Pull back five years and the shape gets clearer.

The IDE stopped being a text editor a while ago. By 2031 it is a control panel for a market of models and agents. When most code is machine-drafted, the human's job is orchestration, judgment, and review.

That is the asymmetric bet. A vendor that owns one great model owns one great model, and models get leapfrogged every few months. Claude Opus 4.7 posted 87.6% on SWE-bench Verified per an April 2026 benchmark summary, while Gemini 3.1 Pro landed at 80.6%. The lead moves. It always moves. Betting on one model is betting on impermanence.

But a vendor that owns the orchestration layer owns the flywheel. Every model that gets better makes the shell more valuable, not less. That is counterpositioning against the closed ecosystems: Microsoft can bundle Copilot tighter, but it cannot easily become model-agnostic without cannibalizing its own bet.

It is unclear whether openness truly wins, or whether the gravity of two or three giant model makers pulls everyone back toward a few defaults anyway. The data is mixed. The 2025 Stack Overflow Developer Survey found trust in AI output fell to about 29% of developers, down from around 40% in 2024, which pushes teams toward controlled, audited pipelines rather than a free-for-all. Openness and governance are on a collision course, and 2031 is where they meet.

My read: the orchestrator posture is the safer long bet. Only shipped features are real. The rest is a roadmap.

What to Build This Weekend

You do not need a CS degree for this. You need one afternoon and a willingness to break things.

First, install LM Studio and download a small open model. LM Studio runs models locally and exposes them through an OpenAI-compatible API, which is the exact door Rider 2026.2 knocks on. Start the local server inside LM Studio and copy the endpoint URL.

Then, in Rider, open AI Assistant settings and point completion at that OpenAI-compatible endpoint. Test it on a real file. Watch your code stay fully local, with zero data egress. That is the "boom" moment, and it costs you nothing per token.

Next, wire up an agent for the heavy lifting. Try Warp, the AI-native terminal that runs shell commands from plain English, so you can compare an agent workflow against local completion. Notice how the two feel different. Cheap and fast for autocomplete, expensive and smart for planning.

Want to go further? Use bolt.new to spin up a full-stack app from a single prompt, then pull it into Rider and let your local model handle the tweaks. First scaffold, then swap engines, then refine. If you are feeding agents outside tools, Apideck's MCP layer gives them permissioned access to 200-plus APIs.

Expect breakage. Local models are slower and dumber than the frontier, and your first endpoint config will probably fail. That is normal. Get your reps in, test aggressively, and learn which model fits which task. The skill worth building is not loyalty to one model. It is knowing how to swap them without blinking.