JetBrains ThinkRail runs on pi, not IntelliJ. 1,100 one-stars say why

Short answer: JetBrains ThinkRail is a new open-source agentic IDE from JetBrains, and it is not built on IntelliJ at all. It is a TypeScript app wrapped around pi, the open coding agent from Mario Zechner and Armin Ronacher, with Microsoft's Monaco editor for tabs. We pulled every rated review of JetBrains' own AI Assistant plugin: 1,100 of 1,704 are one star. Our read: the company that sold the best IDE for 20 years just admitted the harness is the product now, and it did not have one.

TL;DR: 65% of rated reviews of JetBrains AI Assistant on the JetBrains Marketplace give it one star, and its average has not cleared 2.4 stars in any half-year since 2023. pi, the agent ThinkRail embeds, averaged 745,000 npm downloads a day from Sep 16 to Oct 22 2026, more than OpenCode and 3x GitHub's Copilot CLI, and doubled from August to October. A community-built Claude and Codex GUI plugin sits at 4.43 stars with 7% one-stars, so the problem is not AI in the IDE, it is JetBrains' AI in the IDE.

Key numbers

What	Number	Source, checked Oct 25 2026
JetBrains AI Assistant rated reviews that give 1 star	1,100 of 1,704 (65%)	JetBrains Marketplace review API
AI Assistant average rating, all reviews	1.90 stars	JetBrains Marketplace
Community Claude/Codex GUI plugin (CC GUI) average	4.43 stars, 7% one-star (215 reviews)	JetBrains Marketplace
pi npm downloads, average per reported day, Sep 16 to Oct 22	744,752	api.npmjs.org
pi monthly average, Aug vs Oct 2026	391K to 790K a day (+102%)	api.npmjs.org
pi GitHub stars	117,782	github.com/earendil-works/pi
ThinkRail GitHub stars, language	1,720, TypeScript	github.com/JetBrains/thinkrail
ThinkRail on Hacker News	169 points, 115 comments	HN 50233294

JetBrains built its new IDE on someone else's agent

ThinkRail landed on the Hacker News front page on Oct 25 with 169 points. The repo sits under the JetBrains GitHub org with an incubator badge, an Apache 2.0 license and 1,720 stars as of our pull.

Here is the part that should make every IntelliJ loyalist sit up. JetBrains' own blog post, by Maciej Gorywoda on Oct 1, says ThinkRail is "on purpose a thin layer on top of Pi" and that "nothing about how the agent thinks, what it is allowed to do, or how it manages context is ThinkRail's to reinvent. That is Pi's job, and always will be."

Read that again. The company that built the deepest code-understanding engine in the business just said the thinking part belongs to an outside open-source project. ThinkRail embeds pi as a library, runs it in its own process and shares pi's ~/.pi folder, so you can bounce between the pi CLI and the GUI without losing a chat.

And the editor is Microsoft's

The README says the workbench is "a splittable workbench of Monaco editor tabs". Monaco is the editor that powers VS Code. The repo's main language on GitHub is TypeScript, not Kotlin or Java, and the desktop app ships with Electrobun.

So the new JetBrains IDE is a TypeScript shell, a Microsoft editor and an outside open-source agent. That is not a knock on the engineering. It is the clearest signal yet of where JetBrains thinks the value moved.

The receipt: JetBrains' own AI plugin is 65% one-star

We pulled every review on the JetBrains Marketplace for five AI plugins through the public review API on Oct 25: 6,124 reviews, 5,069 with a star rating. Plain version: these are paying and free JetBrains users rating the AI tools inside the IDE they already use.

• JetBrains AI Assistant: 1,704 rated reviews, average 1.90 stars, 1,100 one-stars (65%).
• GitHub Copilot: 2,120 rated, 2.13 stars, 52% one-star.
• Claude Code [Beta] (Anthropic): 488 rated, 2.23 stars, 52% one-star.
• Junie (JetBrains' agent): 542 rated, 2.56 stars, 46% one-star.
• CC GUI (a community Claude and Codex front end by MossX): 215 rated, 4.43 stars, 7% one-star.

The marketplace's own vote tally tells the same story for AI Assistant: 849 of 1,440 votes are one star, 59%. The plugin shows 243,462,459 downloads on its page, which says more about it shipping inside the IDE than about anyone choosing it.

It never got better

AI Assistant's best half-year was its first: 2.39 stars in late 2023. Since then it has sat between 1.69 and 2.00 for six straight half-years, and the 78 reviews written since July average 1.69. That is not a rough launch. That is a product that has been rated the same way for three years while the company kept shipping more AI products next to it.

And it did keep shipping. Count them: AI Assistant, Junie, JetBrains Air (an agent system that works with Claude Agent, Codex, Junie, Copilot and OpenCode) and now ThinkRail. Four AI coding products. One HN commenter, Deukhoofd, asked the obvious question: "What's their strategy here, are they just launching competing products?"

What 1,100 one-star reviews actually complain about

We ran keyword sets over all 1,100 one-star AI Assistant reviews. A review can hit more than one bucket. The biggest bucket is not "the AI is dumb". It is money.

• 360 (33%) talk about price, paying, subscriptions or refunds.
• 180 (16%) talk about quotas, credits, limits or tokens.
• 137 (12%) say it is slow, laggy or freezes the IDE.
• 124 (11%) are about it being bundled, forced or hard to remove.
• 103 mention Copilot, 52 mention Claude and 51 mention Cursor, usually as the thing they would rather use.

The newest one-star on the plugin, from Oct 21, is one sentence: "Without support for a claude subscription plan, this is just useless." One from Oct 20 opens "DO NOT EVER PAY FOR THIS!" in caps. Those two lines are the whole business problem: users already pay for a model, and JetBrains keeps trying to sell them a second meter.

HN commenter dr_faustus put it the same way: their AI strategy "is all over the place and they always wanted to sell their opaque and overpriced tokens, which made no sense."

Follow the money: why pi, and who wins

pi is not a side project anymore. Its GitHub repo has 117,782 stars and 15,087 forks. Its npm package lists Mario Zechner, the libGDX creator, as author, with Armin Ronacher (Flask, Sentry) among the maintainers under the Earendil org.

On npm, which counts every install including CI and bots, pi (both package names combined) averaged 744,752 downloads a day across the 30 days the API reported between Sep 16 and Oct 22. That puts it third among the coding agents we pulled, behind only OpenAI and Anthropic.

Trap we hit: the npm API is missing 7 days in that window (Oct 6, Oct 13 to 15, Oct 19 to 21), shown as zero for every package. We averaged only the days it reported, and the raw daily file is in the download below so you can check us.

The curve that should scare everyone with a closed agent

pi went from zero in October 2025 to 790,000 downloads a day in October 2026. From August to October it doubled, 391K to 790K a day. Over the same two months Claude Code's npm average fell from 2.74M to 1.88M a day, down 31%, while still being the bigger package.

Our bet: this is why JetBrains picked pi and not its own Junie. A minimal, MIT-licensed harness that reads your ~/.claude, ~/.codex, ~/.copilot and ~/.gemini skill folders and lets you plug in any provider is the opposite of a token meter. ThinkRail's README says it straight: "ThinkRail has no accounts of its own." That one line answers 360 one-star reviews.

Who wins: pi and Earendil, who now have the most famous IDE company in the world shipping their engine. Who loses: every vendor whose plan was to own the agent and rent you the tokens. Lowkey, that includes JetBrains' own AI Pro plan.

Hacker News already picked a side

We read and labelled all 35 top-level comments on the ThinkRail thread. Seven liked it or loved the pi core ("I cannot express how happy I am that it uses a Pi core", arikrahman). But 13 of 35 either attacked the strategy and the editor or said they barely open JetBrains anymore.

• tiku: "Jetbrains using vs code is really sad."
• bikelang: "It's an absolute shame that they're using Monaco as their editor for this. Shocking, even."
• theptip: "If I'm in the JetBrains ecosystem isn't it because I prefer their IDEs over VSCode?"
• 1Bad, from the cancel screen: "Best of luck jetbrains I have given you a lot of money figure your shit out and I would happily return."
• throwaw12: "It feels like JetBrains is dying, and they're trying very hard to survive."

Four more spent their comment roasting the README for sounding AI-written, the same slop-police reflex we counted on the Goose language thread. Meanwhile the people who actually use pi were the happiest people in the room.

The roast: JetBrains, you were the moat

For two decades the pitch was simple. VS Code is a text editor with plugins; IntelliJ actually understands your code. That understanding was the moat, and it was real.

Then the agent started doing the understanding. Once the model reads the repo, runs the tests and writes the diff, the IDE's job shrinks to showing you the diff and letting you say yes. ThinkRail's own feature list is worktrees, a chat, a diff view and a spec file. HN user sigotirandolas called new IDEs "mostly just worktrees+sandbox+tabs+prompts".

So JetBrains did the honest thing, and we respect it. It stopped pretending its token meter was the product and borrowed the best open harness it could find. But be clear about what that means: the thing that made JetBrains JetBrains is not in the new JetBrains IDE. It is cooked as a moat, even if it lives on as a product.

Prediction, and nobody has proven it yet: within a year JetBrains either ships pi (or an ACP-compatible open harness) as the default agent inside IntelliJ itself, or ThinkRail quietly becomes the product they push. The 1,100 one-stars are not going to fix themselves.

What this means for you

If you pay for a JetBrains AI plan and a separate model plan, you are paying twice for the same tokens. Check which one you actually use this week, and cancel the other. If you like the IntelliJ editor, keep it for reading code and run your agent next to it; pi, Claude Code, Codex and OpenCode all work from a terminal.

The bigger move: stop marrying the harness. The data says the winners are open, any-provider agents that read skills from plain folders. Keep your rules, specs and skills in files in your repo, not inside one vendor's chat window, and you can switch tools in an afternoon. We track the price side of that on our LLM pricing hub, and every agent fight we cover lives under coding agents. For the outage side of the Claude Code vs Codex fight, see our outage count, and for why token prices keep dropping, the price war piece. More breaking dev stories on /news/.

If you want your agent building what you meant on any harness, including pi or a local model, the $29 AI Builder Kit is five offline apps that keep specs, rules and checks in plain files any agent can read. If it does not get your AI building what you meant, we keep fixing it with you at no extra cost until it does.

Sources

• JetBrains, ThinkRail repo and README, GitHub API stats pulled Oct 25 2026
• Maciej Gorywoda, An agentic IDE that follows Pi's philosophy, Oct 1 2026
• Maciej Gorywoda, Introducing ThinkRail, Aug 19 2026
• Earendil, pi repo and npm package (Mario Zechner, Armin Ronacher), checked Oct 25 2026
• npm downloads API for 17 coding-agent packages, pulled Oct 25 2026
• JetBrains Marketplace reviews for plugins 22282, 26104, 17718, 27310 and 29342, pulled Oct 25 2026
• JetBrains Air product page, checked Oct 25 2026
• Microsoft, Monaco editor, checked Oct 25 2026
• Hacker News, JetBrains Thinkrail thread (169 points, 115 comments), and ThinkRail from JetBrains (56 points), read Oct 25 2026