A terminal coding agent built from scratch in TypeScript. Describe a task in
plain English; acc searches your project, edits files, and runs commands.
Includes permission checks, resumable interactive sessions, automatic context compaction, background commands, and subagents.
Requires Node 22+, macOS or Linux (Windows through WSL), and
ripgrep (rg) on your PATH.
Set DEEPSEEK_API_KEY in your shell before starting. Alternatively, copy
.env.example to .env in your project and fill in the key. ACC uses
DeepSeek V4.1 Flash (deepseek-flash) through the official API, with high
reasoning by default. See model settings.
git clone https://github.com/Xuxyyy/terminal-coding-agent.git
cd terminal-coding-agent
npm install
npm link
cd ~/some-project
accnpm install builds the CLI automatically. ACC is not published on npm.
The workspace is the current directory. Run acc inside your project;
it takes no path argument and refuses to start in your home directory or the
filesystem root. Describe your task at the interactive prompt.
- ACC edits files and runs shell commands. The default
auto-editsmode allows ordinary project edits and asks for approval above its automatic allowance threshold. A saved permission mode takes precedence. - Sandbox defaults to Off. Permission checks still apply, but ACC's OS file
and network restrictions do not. Sandbox On is supported only on macOS;
enable it with
acc --sandbox onor/sandbox. See sandbox details. - Recovery is limited.
/resumereopens interactive sessions./rewindrestores conversation and file-tool backups; it does not undo shell changes. - Subagents run in the background by default. They share your workspace, keep separate conversations, and support messages and saved follow-ups. See subagents.
See commands for session and context controls, and feature status for current capabilities and known limits.
Run one task without an interactive terminal:
acc -p "summarize this project" --max-steps 120 --max-seconds 300Defaults: 120 model turns (including the final answer) and 300 seconds.
Add --json for structured events and a final result. Confirmations are denied
unless --yes is passed; actions already allowed by the permission gate can
still edit files. Print runs do not save resumable sessions.
See print mode for details.
Latest task outcomes as of October 24, 2026, using Harbor and DeepSeek
V4.1 Flash (deepseek-flash, high reasoning):
| Outcome | Tasks |
|---|---|
| Automated verifier pass | 78 |
| Fail | 11 |
| Total attempted | 89 |
78/89 automated verifier passes (87.6%). These latest per-task results
include reruns across multiple ACC source versions, rather than a single
fixed-version pass@1 measurement.
ACC ran with auto-edits, automatic confirmation enabled, and sandbox off.
Some task environments had dependencies prepared before agent execution.
These results describe that evaluation setup; they do not measure interactive
approval behavior or sandbox protection.
See Evaluation for methodology, permission-judge checks, and package/CLI checks.
Documentation lives in docs/.
See CONTRIBUTING.md.
MIT — see LICENSE.