- Book: RAG Pocket Guide
- Also by me: Thinking in Go (2-book series) — Complete Guide to Go Programming + Hexagonal Architecture in Go
- My project: Hermes IDE | GitHub — an IDE for developers who ship with Claude Code and other AI coding tools
- Me: xgabriel.com | GitHub
In November 2022, an AI coding assistant with about half a million users said goodbye. Its founder's farewell post was shared on Hacker News, where it now sits at more than 1,000 points and 528 comments. One line from it became DevClass's headline: "Our 500k developers would not pay to use it."
The company was Kite. On December 7, 2022, less than three weeks after that thread, GitHub announced Copilot for Business at $19 per user per month, a price aimed at the company and not at the developer.
Kite's farewell named two reasons it failed. Four years later, one of them has been overturned and the other has turned out to be a pricing instruction. You can still read the post, though not where it was published: kite.com no longer serves it. A commenter pasted most of the text into the Hacker News thread, other commenters quoted the rest, and the trade press quoted it at the time. Every quote below comes from those copies.
What Kite actually built
Kite was founded in 2014 in San Francisco by Adam Smith, who had earlier founded the Outlook add-on Xobni (DevClass). It raised "tens of millions of dollars in VC backing" (TechCrunch). It took from 2014 to 2019 to reach a market-ready product, and it ran on your own machine, which Kite sold on latency, security and privacy (DevClass).
The product was code completion. On May 12, 2020, Kite launched Kite Pro at $19.90 a month, or $16.60 a month billed yearly, with the paid tier covering Python only at launch (The Register). Smith claimed Kite Pro users were "on average, 18 percent more productive" (DevClass). GIGAZINE's summary of the farewell puts the audience at 500,000 monthly active users, reached with marketing costs close to zero.
For comparison, GitHub Copilot became generally available on June 21, 2022, at $10 a month or $100 a year. Kite had been charging for AI completions for two years by then.
Reason one: the models didn't understand code
The first reason, from the copy of the farewell in the Hacker News thread:
First, we failed to deliver our vision of AI-assisted programming because we were 10+ years too early to market, i.e. the tech is not ready yet.
Then the detail. Kite "fell short of the 10× improvement required to break through because today's state of the art for ML on code is not good enough." And the core of it:
The largest issue is that state-of-the-art models don't understand the structure of code, such as non-local context.
Smith added that "the problem is very engineering intensive," estimated that "it may cost over $100 million to build a production-quality tool capable of synthesizing code reliably," and said of Copilot that "as of late 2022, Copilot has a number of issues preventing it from being widely adopted."
Kite completed the line you were on. The context that mattered usually lived somewhere else in the repo.
Read that in October 2026 and the "not ready yet" part has expired. Claude Code's docs describe a tool that "reads your codebase, edits files, runs commands," and, given a feature request, "writes the code across multiple files, and verifies it works." It "stages changes, writes commit messages, creates branches, and opens pull requests." Codex CLI lets you "inspect code, make changes, run commands, and automate repeatable work without leaving your terminal."
My reading of how non-local context got solved: the model didn't learn to hold your whole repository in its head while you type. The agent goes and gets the context. It opens the file that defines the type, searches for callers, runs the tests and reads the failure. A completion engine gets a few hundred milliseconds and whatever is around your cursor, while an agent can spend minutes in a shell.
Smith was also right that it would take serious money. The models that crossed the line came from labs training general-purpose models, not from a company building one editor plugin.
Reason two: "individual developers do not pay for tools"
The second reason is the one people still quote. Here it is, as quoted in the same thread:
We failed to build a business because our product did not monetize, and it took too long to figure that out.
And the line that made the headline: "Our 500k developers would not pay to use it. Our diagnosis is that individual developers do not pay for tools. Their manager might, but engineering managers only want to pay for discrete new capabilities, i.e. making their developers 18% faster when writing code did not resonate strongly enough." DevClass quotes the "individual developers do not pay for tools" line too.
That paragraph has two claims in it, and 2026 treats them differently.
The first claim, that individual developers don't pay, has not held up as a rule. Anthropic sells Claude Pro at $20 a month billed monthly, and Claude Code is included. The Max plans cost $100 or $200 a month and also include Claude Code. GitHub's individual Copilot plans run from $10 a month for Pro to $39 for Pro+ and $100 for Max. In February 2026, iTWire reported Microsoft's figure of more than 4.7 million paid Copilot subscribers, with Pro+, a tier aimed at individual developers, up 77% in one quarter.
The second claim, that managers pay for discrete new capabilities, aged well. Look at the other half of the same price lists. Copilot Business is "$19 USD per granted seat per month" and Copilot Enterprise is $39 (GitHub docs). Claude's Team plan has a premium seat at $100 a month billed yearly, and Enterprise is $20 a seat with usage billed on top (Claude pricing). Those are bills addressed to a company. Neither vendor's public pricing pages, nor the figures cited here, say how many paid seats a developer buys and how many an employer buys, so nobody outside knows the split. What changed is that both columns now exist.
The unit of work changed
Put Smith's two sentences side by side and the lesson is in the gap between them. "18% faster when writing code" is a speed-up on something the developer was already doing. "Here is a branch with the fix, the tests pass, review it" is a discrete deliverable. A manager has to take an 18% speed-up on faith, but can count a branch with passing tests.
Kite sold keystrokes. The tools people pay for in 2026 sell tasks. Kite itself was reaching that way near the end: the README of its open-sourced codebase lists late 2020 to early 2021 experiments in pull request title generation and code search, and GIGAZINE reports the team considered a pivot to code search before stopping.
Once an agent owns a task, the task has a shape you can name:
# one task, one branch, one folder
git worktree add -b task/fix-login ../fix-login
cd ../fix-login
claude "fix the flaky login test"
# afterwards: what did the task change since it branched?
git diff main...task/fix-login
# land it from the main checkout
cd ~/code/app
git merge task/fix-login
Every step around the agent call is plain git. That's also where the work moves once completion stops being the bottleneck. You're no longer waiting on the model to finish a line. You're deciding which agent is blocked on you, reading what three of them changed, and choosing what's safe to merge.
Where Hermes IDE fits
I build Hermes IDE, so weigh this section with that in mind. Hermes is a desktop app for macOS, Windows and Linux that runs each agent's own interface in a real terminal (Claude Code, Codex, Antigravity CLI, GitHub Copilot CLI, OpenCode and more) and handles the task around it. It is free to download, source-available under BSL 1.1, and not affiliated with Nous Research, Anthropic, OpenAI or any agent vendor.
Press ⌘N (Ctrl+Shift+N on Windows and Linux), type the task, and the agent starts on its own hermes/<name> branch in its own worktree, with the task as its first prompt. That's the git worktree add step from the sketch above, done for you. Agents use their own CLI sign-in, so Hermes never asks for your tokens.
The ⌘N New task sheet in Hermes 2.3.4: Claude Code, the orders-api project, a new hermes/ worktree, Accept edits, model and effort, and a preview of the claude command Hermes will run.
While agents work, every session shows one status: working, needs approval, asked you, done, error or idle. The attention inbox counts the agents blocked on you, and ⌘I jumps to the one that has waited longest. You answer in that agent's own terminal. Hermes never answers a permission prompt for you.
When a task is done, the Review Desk (⌘G) shows the changes by turn or by file, with risk flags, and a line comment goes back to the agent that wrote the line. Land then commits, opens a pull request or merges locally into the branch the task started from, and it can be undone. That covers the git diff and git merge steps from the sketch.
The Review Desk (⌘G) in By turn view: two turns of a Claude Code task in orders-api, turn 1's diff of src/db.ts, the changed files, and a commit message drafted from the turns.
If you only ever run Claude Code, you may not need any of this. Its desktop app already runs sessions side by side and reviews diffs visually. Hermes earns its place when you mix vendors and want one inbox and one review flow across all of them. Hermes 2.3.4 shipped on October 15, 2026.
What the farewell still gets right
Kite open-sourced most of its code under the BSD-3-Clause license (DevClass), and the repository is still up: Go parsers for Python, type induction, a GPT-2-based completion model, years of data pipelines. It's a good read if you want to see how much engineering went into a problem the labs later attacked from a different direction.
The part of the farewell I'd keep is the bar it set. Kite said it needed a 10× improvement to break through, and that 18% did not move a buyer. That test still works on any agent tool you're evaluating, including mine. Does it make the same work a bit faster, or does it hand you a finished task you can review, merge or throw away?
Which one do the tools you pay for today actually do? Tell me in the comments.
If this was useful
Kite's hardest technical problem was non-local context: the code that matters is rarely next to the cursor. Agents now go and fetch that context, and that fetching is retrieval. The RAG Pocket Guide covers chunking, retrieval and reranking for production systems, which is the same problem whether the corpus is your docs or your repo. If you'd rather run Claude Code, Codex and the rest side by side with one review flow, Hermes IDE is a free download and the source is on GitHub.
