Industries / Technology / Software and SaaS / Developer tools
Code editors
The programs in which programmers write and change code, from plain text editors to integrated development environments (IDEs), single programs that also build, run and check the code.
7
Milestones achieved, from 1968 to 2025.
1
To watch for: announced, not yet confirmed outside.
2
Open deliverables. Nobody has shipped them.
The road
AchievedWatch forOpen
Achieved
1968
Searching text by pattern
Butler Lampson and Peter Deutsch wrote the QED text editor for the Berkeley time-sharing system on the SDS 940 computer and described it in Communications of the ACM in December 1967. Ken Thompson, who had used it at Berkeley, wrote a new version for MIT's CTSS time-sharing system. Its main change was regular expressions: a short notation for a pattern of text to find, such as "any line that starts with a digit". Thompson's QED turned each pattern into machine code, a method he published in Communications of the ACM in June 1968 and patented. Dennis Ritchie traces ed and vi, "still much-used on Unix", back to QED.
Credit
Butler Lampson and Peter Deutsch (QED); Ken Thompson (regular expressions in an editor)
1976
Editors that fill the screen
Two editors begun in 1976 are still in use. At Berkeley, the graduate student Bill Joy wrote vi, a "visual" mode that showed a whole screen of text, for the line editor ex that he and Chuck Haley had built from George Coulouris's editor em. Joy said vi's commands are short single keys because he worked from home over a 300 baud modem, a phone-line link so slow that the screen redrew more slowly than he could think. At MIT, Guy Steele and Richard Stallman built Emacs as a set of macros for the TECO editor. Its commands were written in a programming language, so users could add new commands while editing. In Stack Overflow's 2024 survey of developers, Neovim, a modern descendant of vi, was the most admired editor: 83% of its users wanted to keep using it.
Credit
Bill Joy and Chuck Haley (vi); Guy Steele and Richard Stallman (Emacs)
1983
Editor and compiler together
On 20 November 1983, by the account of Embarcadero, which now sells its successor Delphi, Borland released Turbo Pascal: a text editor and a compiler, the program that turns code into instructions the computer runs, in one program held in memory. Before it, a typing mistake on an IBM PC or a CP/M machine could mean minutes of compiler and linker passes over floppy disks; with Turbo Pascal the cycle took moments. It sold by mail for $49.95 when rival compilers cost hundreds of dollars. The Register credits Philippe Kahn, Borland's chief, with the idea of the all-in-one tool.
Credit
Anders Hejlsberg (compiler) and Philippe Kahn, Borland
2015
A free editor takes over
In November 2015 Microsoft made its free code editor, Visual Studio Code (VS Code), open source, and added extensions: add-on programs that anyone can write and share through a public marketplace. VS Code is built on Electron, a framework that came out of GitHub's Atom editor. In Stack Overflow's 2024 developer survey, 74% of respondents had used VS Code in the past year, more than twice as many as used its nearest alternative, Visual Studio.
Credit
Microsoft
2016
One protocol, many languages
Features such as autocomplete, "go to definition" and error warnings had to be written again for each editor and each programming language. On 27 June 2016 Microsoft, Red Hat and Codenvy announced that the Language Server Protocol (LSP) was being adopted across the industry. With LSP, the knowledge of a language lives in a separate program, a language server, which talks to the editor through a shared set of messages, so one server can serve any editor that speaks the protocol. Outside adoption followed: Emacs 29.1, released on 30 July 2023, ships Eglot, a client for the protocol.
Credit
Microsoft, with Red Hat and Codenvy
2022
An editor switched off
GitHub introduced the Atom editor in 2011. On 8 June 2022 it said it would archive Atom on 15 December 2022: Atom had had no significant feature work for several years, community involvement had "declined significantly", and GitHub wanted to focus on VS Code and its cloud editor, Codespaces. After the sunset Atom's package manager, which installed its add-ons, stopped working. In January 2023 GitHub revoked Atom's code-signing certificates after a break-in, and some versions stopped running on 2 February 2023. Atom's makers went on to build Zed, which reached version 1.0 on 29 April 2026.
LessonAn editor owned by one company lasts only as long as it serves that company, and its users' add-ons can stop working the day it ends.
Credit
GitHub
2025
A new IDE withdrawn
JetBrains introduced Fleet in 2021 as an IDE "built from scratch", lighter than its IntelliJ-based tools, which date from 2001. Fleet never left public preview, and from 22 December 2025 it was no longer offered for download. "We could neither replace IntelliJ IDEA with Fleet nor narrow it into a clear, differentiated niche," JetBrains said. DevClass reported that most developers stayed with the IntelliJ tools and their plug-ins. Fleet's code now runs Air, a JetBrains tool in which, the company says, the developer guides an AI agent and reviews its output.
LessonDevelopers rarely leave an editor whose plug-ins and habits they already have, so a new one needs work the old one cannot do.
Credit
JetBrains, as reported by DevClass
Watch for
Any agent in any editor
The Agent Client Protocol (ACP), which grew from work by Zed and Google to bring Google's Gemini CLI agent into an editor, is meant to do for AI coding agents what LSP did for languages: an agent built once works "in any supporting editor". JetBrains and Zed opened a shared registry from which agents can be installed inside their editors; JetBrains names Gemini CLI, Claude Code, Auggie, OpenCode and Copilot among the most common agents.
Achieved whenVS Code, the most used editor, supports ACP itself, or an outside test runs the same registry agents in at least three editors and publishes which features work in each.
Claimed by
JetBrains and Zed, January 2026
Open
01
Keystroke delay, measured alike
A public table of typing latency, the time from pressing a key to seeing the character on screen, for ten editors: VS Code, Cursor, Zed, Visual Studio, IntelliJ IDEA, Sublime Text, Neovim and Emacs in a terminal, Emacs in its own window, and Notepad as a baseline. Each is timed in a small file and in a 50,000-line project, with AI features on and off, on Windows, macOS and Linux. A small board plays the keyboard and a light sensor on the screen sees the character appear, so no editor's own clock is trusted.
The briefCredit · unclaimed02
What AI editors upload
A public test of what code editors with AI features send off the machine. A dummy project holds marked "canary" strings: in its source files, in a.env file of fake passwords and keys, in a file the editor's ignore list says to skip, and in old git history. Each editor does the same five tasks: explain a file, fix a planted bug, write a test, let its agent do a refactor, and sit idle for 10 minutes. A proxy records the traffic and a script searches it for the canaries. Each run is done with the editor's privacy setting on and off.
The briefCredit · unclaimed
Open 01 · the brief
Keystroke delay, measured alike
The first to publish is credited on this line, and the deliverable moves from Open to Achieved.
What to build
A public table of typing latency, the time from pressing a key to seeing the character on screen, for ten editors: VS Code, Cursor, Zed, Visual Studio, IntelliJ IDEA, Sublime Text, Neovim and Emacs in a terminal, Emacs in its own window, and Notepad as a baseline. Each is timed in a small file and in a 50,000-line project, with AI features on and off, on Windows, macOS and Linux. A small board plays the keyboard and a light sensor on the screen sees the character appear, so no editor's own clock is trusted.
Why it is still open
Pavel Fatin measured popular editors in 2015 with Typometer, a tool he wrote that reads the screen to time each character; his tests predate today's AI editors. A 2026 blog test by Vantaige compared only Zed, Cursor and VS Code on one MacBook, took the median of three runs, and reports about 2 ms for Zed and about 12 ms for the other two without saying how it measured input. A search found no test that covers many editors on several systems with the AI features switched on.
Built with what exists
Typometer for a software-only cross-check; a microcontroller board that can act as a USB keyboard, such as an Arduino Leonardo; a photodiode, a light sensor, read by the same board; a monitor running at 120 Hz or more; Python for the statistics.
First step
Build the board and sensor, then time 200 keystrokes in Notepad and in VS Code and compare the results with Typometer's.
Done looks like
10 editors, 3 systems, 2 file sizes, AI on and off, 200 keystrokes in each condition, with the median and the 95th percentile in milliseconds, raw timings and scripts published. A second person with different hardware reruns three editors and gets the same order, with medians within 3 ms of the first run.
Open 02 · the brief
What AI editors upload
The first to publish is credited on this line, and the deliverable moves from Open to Achieved.
What to build
A public test of what code editors with AI features send off the machine. A dummy project holds marked "canary" strings: in its source files, in a.env file of fake passwords and keys, in a file the editor's ignore list says to skip, and in old git history. Each editor does the same five tasks: explain a file, fix a planted bug, write a test, let its agent do a refactor, and sit idle for 10 minutes. A proxy records the traffic and a script searches it for the canaries. Each run is done with the editor's privacy setting on and off.
Why it is still open
The closest work maps traffic, not leaks. Chaser Systems, a network security firm, captured the traffic of seven coding agents and editors, among them Cursor, Zed, Copilot in VS Code and Claude Code, on a fixed set of tasks between July and September 2025. It recorded which internet domains each contacted and how much data went where with telemetry on and off, and whether each would read cloud credentials when asked; it planted no marked strings in files the tools were told to skip. Other tools have been examined one at a time: a Cereblab researcher captured the traffic of xAI's Grok Build command-line agent and found.env secrets sent verbatim and 5.10 GiB uploaded from a 12 GB repository, as Developers Digest reported in July 2026, and Knostic found in December 2025 that Claude Code loaded.env files from the project folder automatically. A search found no repeated test that tracks planted strings across the editors developers use most.
Built with what exists
mitmproxy, an open-source proxy that can show the encrypted traffic of programs set to trust it; VS Code with GitHub Copilot (see the Note "Effort, not model"), Cursor, Zed, Windsurf and a JetBrains IDE with its AI Assistant; ripgrep to search the captures; fake secrets only, in a project made for the test.
First step
Set up mitmproxy with one editor, plant five canaries in a sample project, run one task, and search the capture.
Done looks like
6 editors, 5 tasks, privacy on and off: 60 runs, each listing which canaries left the machine and to which internet domains, with the captures published. Each maker gets 30 days' notice before publication. A second person reruns 10 runs picked at random and finds the same canaries.