Kite

Kite is discontinued. Founder Adam Smith published "Kite is saying farewell" saying that "From 2014 to 2021, Kite was a startup using AI to help developers write code. We have stopped working on Kite, and are no longer supporting the Kite software." TechCrunch reported on 10 December 2022 that Kite "abruptly shut down last month", which puts the public announcement in November 2022.

kite.com is now a single static page. Fetching /, /about/, /blog/, /download/, /python/docs/ and /robots.txt all return the same 58,928-byte document — a SingleFile-saved copy of the farewell post whose own save header reads "saved date: Sat Mar 29 2025". The footer still links to Kite Team Server, JavaScript Completions, Copilot, Linux Support and thirteen editor-integration pages; none of them exist as separate pages any more.

The product infrastructure is gone, not merely idle. linux.kite.com, windows.kite.com, release.kite.com, ga.kite.com and help.kite.com all fail DNS resolution. That is the practical difference between "unmaintained" and "unusable": there is no host left to download the Kite Engine from.

One thing that is not gone is the domain registration. Verisign RDAP shows kite.com registered on 18 January 1995, last changed 16 May 2026, expiring 19 January 2035, on Cloudflare nameservers via Amazon Registrar. Somebody is paying to keep it alive a decade out while advertising that it may be for sale.

Above the farewell post the page now carries a yellow banner reading "kite.com may be for sale", with an offers link to a Google Form and a pull-quote from a16z's Martin Casado about brand effects. There is no asking price on the page.

Wayback snapshots bracket when this went up: the 22 May 2025 and 1 April 2025 captures contain no "for sale" text; the 1 July 2025 capture does, as do 6 August 2025 and 4 April 2026. So the banner has been in place for at least thirteen months and appeared somewhere between late May and early July 2025.

The practical reading: whoever controls kite.com is treating it as a domain asset, not a dormant product. If you are considering anything that depends on kite.com URLs staying put — a docs link, a citation, a redirect — assume the address can change hands.

It is worth being specific about this, because the parts Kite got right are the parts still worth reading.

On-device inference, early. With the January 2019 funding round Kite shipped models that ran locally, without an internet connection, across plugins for Atom, Sublime Text 3, VS Code, Vim, PyCharm and IntelliJ. TechCrunch reported this at the time as both a privacy answer and a deliberate way to avoid running cloud inference infrastructure. Local completion models were not the norm in 2019.

A purpose-built Python analysis pipeline, not a wrapper. Per the codebase README, Kite crawled GitHub (about 20 TB), then ran package exploration that imports and introspects Python packages, statistical type induction that "estimates the return types of functions based on attributes accessed on their return values across GitHub", dynamic analysis that executed 2–3k short scripts to extract runtime types, and return-type extraction from docs — unifying those sources into per-package resources (symbol graph, arg specs, popular call signatures, kwargs, return types). The README's own example: type induction put a 51.9% probability on numpy.matrix.argmin returning a numpy.matrix.

Their own Python parser in Go, robust to syntax errors — it can parse partial function calls, which is exactly the state a file is in while you are typing. That is the unglamorous engineering that makes completion work.

Distribution. Kite says it reached 500,000 monthly-active developers "with almost zero marketing spend". That figure is the founder's own and we have not seen it independently audited, though TechCrunch repeats it as Smith's account; for scale, TechCrunch reported "about 30,000 Python developers" in January 2019.

Candour on the way out. The farewell post and the accompanying code release are more specific about what went wrong than most shutdown notices manage. That is a contribution in itself, and it is why the next section exists.

This is the only part of Kite you can still use.

github.com/kiteco/kiteco-public is the primary repo, licensed BSD-3-Clause, 769 stars and 174 forks, about 47 MB, last pushed 1 December 2022 (all figures from the GitHub API, checked 15 August 2026). The wider kiteco org has 40 public repos and 2,084 followers, including the editor plugins (vscode-plugin, 923 stars; intellij-plugin; vim-plugin; KiteSublime; jupyterlab-kite) and the old issue-tracker, which at 1,737 stars and 413 open issues is the most-starred thing in the org.

Two caveats stated by the repo itself. First, its own description is "Primary Kite repo — private bits replaced with XXXXXXX", and the README says: "It has been lightly adapted for publication here by replacing private information with XXXXXXX. As a result many components here may not work out of the box." Second, the getting-started instructions are the internal ones — they assume Go 1.15.3, Git LFS, Travis CI, VPN credentials from a named colleague and SSH keys from Kite's Dropbox. Treat this as source to read and lift from, not a project to build.

The README also offers the derived data — "You can download the 800MB zip file with all the Python open-source packages here" — but that Google Drive link returns 404 on both the /file/d/.../view and uc?export=download endpoints. The author's own hedge in the README turned out to be right: "please let me know soon if you want anything because we will likely end up archiving all of this." A repo of readme_assets (per-package examples, product specs for lexical and multi-provider completions) is still in the repo.

An exact mirror is on SourceForge as project kite.mirror if GitHub is blocked for you.

Who this is actually for: someone building Python code intelligence who wants a worked example of GitHub-scale package introspection and statistical type inference in Go, or someone writing about the pre-LLM era of code completion. The README's Q&A on deployment, model evaluation, parsers and why they avoided ML linting is the densest part and is readable in ten minutes.

The install paths are in three different states, which is confusing enough to be worth spelling out.

VS Code Marketplace: delisted. A gallery query for kiteco.kite returns TotalCount 0, and a text search for "kite autocomplete" surfaces no Kite extension.

Open VSX: still listed, still downloadable, and will not work. kiteco/kite version 0.147.0 is there, published 3 June 2021, 41,435 downloads. The extension is only a client — it drives a separately-installed Kite Engine desktop app, and the hosts that distributed that engine no longer resolve. Installing it gets you an extension that cannot complete its own setup.

JetBrains Marketplace: gone, and the name is now taken by something else. There is no Kite AI plugin. Searching the marketplace for "kite" returns a plugin called Kite (id 29192) from vendor KiteLang — an Infrastructure-as-Code language positioned as a Terraform alternative, entirely unrelated to this Kite.

The npm package kite-installer is still published, latest 3.16.1, last modified 19 June 2022 — also a client for infrastructure that is gone.

The name is heavily reused and several of the others are live products, so it is easy to arrive at this page by mistake or leave with the wrong impression:

• Kite by Zerodha (kite.zerodha.com) — the Indian brokerage's trading platform, live and unrelated. Its Kite Connect API is what most "kite" GitHub topics refer to.
• KiteLang's Kite — an Infrastructure-as-Code language, live as a JetBrains plugin, unrelated.
• Kite by Kagi (kagisearch/kite-public) — Kagi's news app, unrelated.
• Kite AI / gokite.ai — a live site under a different domain, unrelated to this company.

If a search result or an AI assistant tells you "Kite" is active, check which one it means before concluding this listing is out of date.

Smith's post-mortem is unusually specific, and the failure modes are ones any developer-tools buyer or builder will recognise.

On the technology: "we failed to deliver our vision of AI-assisted programming because we were 10+ years too early to market". He judged the state of the art insufficient — "state-of-the-art models don't understand the structure of code, such as non-local context" — and estimated "it may cost over $100 million to build a production-quality tool capable of synthesizing code reliably, and nobody has tried that quite yet." This was written with GitHub Copilot already shipping, which he assessed as showing "a lot of promise but still has a long way to go."

On the business, which is the part that has aged best: "Our 500k developers would not pay to use it... 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." They reached product-market fit in 2019, five years in; found a pivot direction in code search; and did not have the energy left to run it.

Company and funding: Kite was San Francisco-based, founded by Adam Smith, with the GitHub org created in April 2014. It raised a $17M round announced 28 January 2019, led by Trinity Ventures with personal participation from Nat Friedman, then about to become GitHub's CEO. Smith told TechCrunch that $7M of prior funding converted into stock as part of that total. TechCrunch's 2022 piece describes the total backing as "tens of millions of dollars".

Included because it is part of Kite's record in the Python community and a reader researching the name will find it.

In 2017 The Register reported that Kite employees had, without disclosing the affiliation, added Kite-promoting code to two popular Atom packages: autocomplete-python, where an installer presented a choice between Kite and an unnamed "local engine" (in fact the Jedi library), and minimap, which gained promotional links. Developers called the selection UI a dark pattern; minimap was forked, and Kite's engineer subsequently apologised and said the changes would be undone — "The next release will no longer show anything."

Jedi's maintainer David Halter, the party most directly affected, was measured about it: "I have talked to them and they seem to be decent guys. They just don't understand that code in the cloud is something very problematic", while also saying the option "never should have been added" and that the backlash was "not entirely undeserved". Kite later moved inference on-device, which addressed the underlying cloud-upload objection.