Best open-source alternatives to Claude Code - CLI

OpenCode is an open source AI coding agent designed for developers who want to work from the terminal. It provides a TUI-focused workflow for coding, planning, and code exploration, and the README emphasizes that it is provider-agnostic rather than tied to a single model vendor. The project is aimed at people who want an agentic coding assistant with flexibility in deployment and model choice. It includes built-in agents for full-access development work and read-only analysis, plus a general subagent for complex searches. The README also notes a desktop app beta and a client/server architecture, which suggests the tool can be used locally while being controlled remotely from other clients.

Crush is a terminal-based coding assistant for developers who want to use large language models from inside their existing shell workflow. It is designed to connect a user’s codebase, tools, and contextual signals to a chosen LLM provider, and it supports a broad range of model backends including OpenAI-compatible and Anthropic-compatible APIs. The project emphasizes session-based work, letting users maintain multiple contexts per project and switch models during a session without losing context. It also integrates with language servers for richer code awareness and supports Model Context Protocol servers for extending its capabilities. Crush is distributed across many platforms and is available through package managers, native binaries, and source installation, making it suitable for users who want a configurable, cross-platform AI coding tool in the terminal.

Flue is a sandbox agent framework designed for developers building autonomous AI agents. It presents itself as a way to structure agents around secure, isolated sandboxes, combining an agent harness with either a virtual sandbox or a full container environment. The README frames the project as suitable for anyone already familiar with tools like Claude Code, OpenCode, Codex, or Gemini. The project appears aimed at building several kinds of agents: quick API-driven agents, support agents backed by filesystem-mounted knowledge bases, CI triage agents with controlled command access, and full coding agents that operate in cloned repositories. It uses TypeScript packages, supports webhook triggers, and integrates with external runtime components such as Cloudflare R2, Daytona, and remote MCP servers. The examples emphasize structured outputs, role-based subagents, and sandbox-aware tool access.

Orca is a desktop application for coordinating AI coding agents across multiple repositories and worktrees. It is aimed at builders who want to run several CLI-based agents in parallel while keeping each agent’s work isolated, visible, and easy to manage from one interface. The project emphasizes workflow orchestration rather than model hosting: users bring their own subscriptions or agent tools, and Orca provides the surrounding workspace for tabs, panes, diffs, commits, SSH connections, GitHub-linked context, and notifications. The README also highlights companion mobile control and several feature showcase pages for workflows such as parallel worktrees, terminal splits, design mode, and remote SSH worktrees.

Dirac is an open-source AI coding agent designed for precision refactoring and code modification with a strong emphasis on token efficiency. It is intended for developers who want an assistant that can read, edit, and reason about codebases while minimizing context waste and reducing API costs. The project supports both a VS Code extension and a command-line interface. Its workflow centers on curated context, hash-anchored edits, AST-aware structural changes, and multi-file batching so it can carry out complex changes across repositories with fewer model calls. The README also highlights an approval-based workflow, common task commands, and support for multiple AI providers through environment variables or interactive authentication.

zerostack is a lightweight coding agent written in Rust for terminal-based development workflows. It is aimed at developers who want an interactive agent that can work with code, run tools, manage sessions, and operate across different model providers without the overhead of larger JavaScript-based alternatives. The project emphasizes speed, small resource usage, and configurable control. It offers a terminal UI, multiple prompt modes for different tasks, session persistence, a permission system with several approval levels, and experimental features for looping on long tasks and working across git worktrees. It also supports multiple AI providers and custom providers through configuration, with optional sandboxing via bubblewrap for safer command execution.