v0.1.13 · C11 · zero dependencies

A coding agent that does
the work in your terminal

Ask in your own words: read code, edit files, run commands, handle git, search the web. Delta executes it — in your project, with the model you choose. And it remembers the context next time.

Download for your system See how it thinks

Windows · Linux · macOS (source) — one file, no Node, no Python, no admin

636 KBa single executable
for everything
0runtime
dependencies
40+tool operations
day to day
3systems with their own
installer

overview

One binary. One agent.

Delta talks to the model and does the actual work: opens the file, edits the block, runs the test, reads the error, fixes it. All in the terminal, with the same mental model as an editor.

01

It executes, not just answers

Reads and edits files by exact blocks, runs commands with timeouts, keeps background jobs, shows diffs and commits. The result appears in the chat, not in another tab.

02

It remembers

Sessions per folder in JSONL: delta starts a new chat, delta continue reopens that folder's last one with the full history and the model's thinking. Skills become SKILL.md; durable facts go to memory.

03

You steer mid-turn

Type while it works. The message joins the queue and it cuts the reply at the next chunk to read it — no waiting for the turn to end.

how it thinks

The turn loop

The agent owns the loop; the model only asks. Every tool request becomes a real execution in your project and the result goes back so the model can decide the next step.

01You write the request
02The model streams a reply
03It asks for tools (JSON)
04Delta runs and returns
05Repeat until done

Incremental render

Each block remembers the line where it starts; the transcript is rebuilt from the first dirty block. A typical frame is 0,5 ms even with hundreds of blocks.

Native terminal

Raw mode with ReadConsoleInput on Windows and termios on POSIX: mouse, paste, resize, alternate screen and modified-key shortcuts.

Lean network

WinHTTP on Windows, curl on POSIX. SSE read line by line, aborting immediately when you step in mid-stream.

what it does

One agent. No helpers needed.

Local tools, no service in between: your project, your shell, your git.

Files

  • read with offset and limit
  • edit by exact blocks
  • create, delete, move
  • search and list structure

Shell

  • commands with timeout
  • background jobs with logs
  • /terminal <cmd> inside the chat
  • UTF-8 output on Windows

Git

  • status, diff, log
  • commit with a message
  • checkpoint before changes
  • roll back safely

Web

  • download pages, follow redirects
  • extract links and data
  • session with cookies
  • reader mode

Skills & MCP

  • workflows in SKILL.md
  • indexed in the system prompt
  • MCP servers (JSON-RPC)
  • HTTP or stdio transport

Memory & graph

  • durable facts by scope
  • text search
  • links between notes
  • project graph of symbols/imports

why delta

Built to stay
out of the way

One file, no runtime

636 KB of C11. Opens instantly and does not depend on Node, Python or a framework upgrade.

Your data in your folder

Keys, history, memory and skills live in .deltaCore. Nothing is scattered across the system.

You choose the model

models.dev catalog: hundreds of models. Switch with /models and tune effort with /thinking.

No black box

Every tool shows its arguments, output and timing in the chat. You review what it did before trusting it.

install

From zero to your first request in minutes

Every download is the installer: it asks for the folder, copies the binary, puts delta on your PATH and creates the shortcut.

Windows 10/11 x64 tested

Delta.exe — program + installer in one file. Double click opens the HTML install window (no console, native folder picker). This is the reference platform.

Linux x86_64 builds & runs

delta-linux-x86_64.run — self-extracting, installs and puts delta on PATH. Verified in WSL; the desktop bits (.desktop, GNOME favorites, clipboard) were not exercised on a real desktop.

macOS source not tested

Source package + INSTALAR-MAC.command: builds on your machine. I have no Mac, so the native parts (osascript, .app, Dock) were never run — treat it as a starting point.

# linux
chmod +x delta-linux-x86_64.run && ./delta-linux-x86_64.run

# first run
delta                    # opens the CLI (always a new chat)
delta continue           # reopens this folder's last chat
/login                   # pick the provider and paste the key
/projects D:\dev\app     # where it works (saved)

# day to day
read src/main.c and explain the flow
/terminal git status     # runs in your shell and shows the output in the chat
delta update             # installs the newest release
delta wiki               # prints the documentation address

Commands

  • /models connected models
  • /thinking effort
  • /verbosity how much prose
  • /undo restore a file
  • /subagentmodel model for subagents

Shortcuts

  • Enter send · Shift+Enter new line
  • Ctrl+C stop (never closes) · with a selection: copy
  • Ctrl+V paste text/image
  • drag the mouse to select · click clears
  • Ctrl+Q return the queue to the editor

Data

  • config.json · auth.json
  • sessions/ per folder
  • skills/ · mcps/
  • system/prompt.md your rules

faq

Short answers

Do I need an API key?

Yes, from a provider you use (OpenAI, Anthropic, OpenRouter, DeepSeek, opencode…). /login asks for the key once and stores it per provider in auth.json.

Does it run offline?

The program runs; the model does not. Without network it still opens the CLI, sessions and local tools — but it cannot talk to the provider.

Why C11?

A single executable with no runtime opens instantly and never breaks on a Node upgrade. The TUI, JSON parser, markdown and HTTP client are all written in the project.

Windows said the app is unknown

That is SmartScreen because the binary is not code-signed: More info → Run anyway.

What about macOS?

The binary has to be built on the machine — there is no reliable cross-compile to Mach-O. The package ships a .command that checks Xcode Command Line Tools, builds and installs. It was never run by the author (no Mac available), so the macOS path is untested: expect to fix things, not to have a supported binary.

Was Linux tested on a real desktop?

Partially. The build, the CLI, the TUI, real streaming and the self-update were verified on Linux (WSL, gcc 13). The desktop integration — .desktop entry, GNOME favorites and the clipboard (wl-copy/xclip) — was not exercised on a desktop session.

Can I use it commercially or publish a fork?

No. The license is source-available and non-commercial: it forbids selling, charging for access and cloning (public fork, package, binary or rebrand). Sharing the unmodified copy, non-profit and with credit, is allowed. See LICENSE.

Where does my data go?

Only to the provider you configure. History, keys, memory and skills stay in your .deltaCore folder.