AI/TLDR

Dirac

A long-running coding agent for VS Code, the terminal and ACP editors

Autonomous Coding AgentsOpen source
Updated
28 Apr 2026
Language
TypeScript
License
Apache-2.0
Coverage
1 story
$npm install -g dirac-cli

What's new

28 Apr 2026

AI/TLDR covered Dirac's arrival as an open-source coding agent: an Apache-2.0 fork of Cline that runs in VS Code, a terminal CLI and ACP-compatible editors, with hash-anchored editing, AST inspection and a walk-away Goal mode.

Latest news

Overview

Dirac is an open-source coding agent built for long-running software-engineering work rather than one-shot edits. It runs in VS Code and Open VSX editors, from a terminal CLI, or as an external agent inside any client that speaks the Agent Client Protocol (ACP), including JetBrains IDEs and Zed. Provider configuration is stored by Dirac itself, so the same credentials and model choices follow you between those interfaces.

The project's focus is efficiency: instead of brittle search-and-replace blocks, Dirac edits files through a stable line-anchor protocol so a model can name an exact source range and replace only that range even after nearby lines shift. It also reads and rewrites code structurally through syntax trees — `inspect_ast` can return an outline of many files or one exact implementation, and `edit_ast` can replace a single function or rename indexed references across a directory — which keeps whole files out of the context window.

Dirac is a fork of Cline released under the Apache License 2.0, built by Dirac Delta Labs. It supports dozens of providers and hundreds of models through API keys, subscription-backed accounts, cloud platforms such as AWS Bedrock and Vertex AI, and any OpenAI-compatible endpoint, so the model and the credentials are yours to choose.

What it does

  • Goal mode: give Dirac an objective with `/goal` and it creates and coordinates its own tasks toward it, pausing when it needs your input
  • Hash-anchored file editing — a stable line-anchor protocol that replaces an exact source range instead of matching text, reducing edit payloads and retries
  • AST inspection and manipulation: outline files or fetch one implementation, replace a function by symbol, or rename indexed references in a single call
  • A separate cheaper Utility model handles context compaction, task handoffs, commit messages and first-pass permission decisions
  • Self-extension with `/new-tool <description>`: Dirac writes, compiles, validates and smoke-tests a typed tool you can use in the same conversation
  • Continuous steering — send a message mid-task and it is delivered with the next tool response instead of cancelling the run

Getting started

Dirac ships as a VS Code / Open VSX extension and as an npm-installed CLI. The CLI requires Node.js 22.13 through 24.x; Node.js 25 is not supported.

Install the CLI

Install the package globally, then authenticate against the provider you want to use.

bashbash
npm install -g dirac-cli
dirac auth

Run your first task

Pass a prompt directly, or start the interactive composer with no arguments.

bashbash
dirac "Analyze the architecture of this project"

Other ways to start a run

Plan mode drafts an approach first, --yolo runs unattended with plain output, and you can pipe context straight in.

bashbash
dirac --plan "Design this feature"
dirac --yolo "Fix the tests"
git diff | dirac "Review this"
dirac history

Use it in an editor instead

Install Dirac from the Visual Studio Marketplace or Open VSX and open its sidebar, or add it from your editor's ACP Registry (JetBrains 2025.3+ under Settings → Tools → AI Assistant → Agents, or Zed under Agent Settings → External Agents). You can also run `dirac --acp` manually.

Commands and code are distilled from the project's own documentation — always check the official repo for the latest.

When to use it

  • Hand an agent a multi-hour refactor and let Goal mode keep working while you check in, steer or pause it
  • Make precise structural changes — rename a symbol across an indexed codebase, or replace one function — without the agent reading whole files
  • Work from whichever surface you are already in: VS Code, a terminal, or a JetBrains/Zed session over ACP, with one provider configuration
  • Cut token spend on routine agent overhead by routing compaction, handoffs and permission checks to a cheaper Utility model

How Dirac compares

Dirac alongside other open-source autonomous coding agents tools AI/TLDR tracks, ranked by GitHub stars.

ToolStarsWhat it does
opencode★ 208kOpenCode is an open source AI coding agent that runs in your terminal, with built-in build and plan agents and an optional desktop app.
Claude Code★ 146kAnthropic's agentic coding tool for the terminal, IDE and GitHub: it understands your codebase, executes routine tasks, explains code and handles git workflows from natural-language commands.
OpenAI Codex★ 125kOpenAI's cloud and CLI coding agent that writes features, fixes bugs, and proposes code changes across a repo, running tasks in parallel.
Gemini CLI★ 107kAn open-source command-line AI agent from Google that connects your terminal to Gemini models for reading code, editing files, running shell commands, and searching the web.
Pi★ 107kMinimal terminal coding agent harness with four built-in tools, extended through TypeScript extensions, skills and prompt templates instead of forks.
autoresearch★ 96.3kKarpathy's minimal harness that hands a coding agent a single-GPU LLM training script and lets it run fixed five-minute experiments, keeping or discarding each change on its own.
OpenHands★ 88.4kAn open-source AI software-development agent that plans tasks, edits files, runs commands, and tests code, usable from a terminal CLI, a local web GUI, or a Python SDK.
Dirac★ 1.5kA long-running coding agent for VS Code, the terminal and ACP editors