Open-source terminal agent that works with any model provider.
A TUI coding agent with LSP awareness, multi-session work, and a provider-agnostic model layer, so you can mix hosted APIs and local models. It runs on your machine and edits your working tree directly.
- Best for
- Developers who want a Claude Code-style agent without vendor lock-in
- Pricing
- Free; you pay the model provider
- Platforms
- macOS, Linux, Windows
- Licence
- MIT
Client/server design also powers editor integrations; model choice changes quality a lot.
Anthropic's terminal agent for repository-wide changes.
Reads the project, plans multi-file edits, runs commands and tests, and can commit or open pull requests. Works interactively, in a headless mode for scripts, and inside IDEs.
- Best for
- Teams already using Claude models and long agentic sessions
- Pricing
- Claude Pro/Max subscription or API usage
- Platforms
- macOS, Linux, Windows
- Licence
- Proprietary
The GitHub repository tracks issues and docs; the agent itself is closed source.
Open-source local coding agent from OpenAI.
A lightweight terminal agent that can read, patch and run code in a sandbox, with approval modes for how much it may do on its own. Also available as a cloud task runner.
- Best for
- Local, sandboxed automation with GPT models
- Pricing
- Included with ChatGPT plans or API usage
- Platforms
- macOS, Linux
- Licence
- Apache-2.0
Sandbox and approval settings matter more than the prompt for safe use.
Google's open-source agent with a generous free tier.
Terminal agent built around Gemini models, with a very large context window, MCP server support and scriptable non-interactive mode. Free quota with a personal Google account.
- Best for
- Large-repository exploration on a budget
- Pricing
- Free tier; paid API for heavier use
- Platforms
- macOS, Linux, Windows
- Licence
- Apache-2.0
Free limits change often; check current quotas before relying on them.
CLI agent tuned for Qwen coding models.
A terminal agent derived from Gemini CLI and adapted to Qwen3-Coder, with free daily quota through Qwen's hosted API. Supports local OpenAI-compatible endpoints.
- Best for
- Cheap coding agents in Chinese or English
- Pricing
- Free quota; pay-per-token beyond it
- Platforms
- macOS, Linux, Windows
- Licence
- Apache-2.0
Works best with Qwen models; other providers need extra configuration.
Git-native pair programming in the terminal.
Creates a repository map, proposes edits, and writes clean commits for every change. Supports dozens of models and a watch mode that reacts to file changes.
- Best for
- Developers who want small, reviewable commits
- Pricing
- Free; pay the model provider
- Platforms
- macOS, Linux, Windows
- Licence
- Apache-2.0
Best with strong code models; long sessions need context management.
Extensible on-machine agent from Block.
Runs locally as a CLI or desktop app, connects to MCP extensions for tools and data, and can automate engineering chores beyond code editing.
- Best for
- Tool-heavy workflows and local automation
- Pricing
- Free; you pay the model provider
- Platforms
- macOS, Linux, Windows
- Licence
- Apache-2.0
Extension quality varies; start with the built-in recipes.
Terminal agent for large, multi-file tasks.
Plans a change, applies it in a sandboxed copy, and lets you review diffs before touching your repo. Designed for tasks too big for a single prompt.
- Best for
- Big refactors with a review step
- Pricing
- Free self-host; cloud plans
- Platforms
- macOS, Linux
- Licence
- AGPL-3.0
More setup than single-purpose agents.
Open platform for software engineering agents.
Formerly OpenDevin. Gives an agent a browser, terminal and editor in a sandbox, available as a CLI, local web UI or hosted cloud.
- Best for
- Running agents in an isolated environment
- Pricing
- Free self-host; cloud usage-based
- Platforms
- Docker, Linux, Web
- Licence
- MIT
Docker is the normal install path; resource use is higher than CLI agents.
Research agent that fixes GitHub issues.
A minimal agent loop that turns issues into patches, used as a benchmark baseline by research teams. Configurable tools and model backends.
- Best for
- Evaluating agent performance on real issues
- Pricing
- Free; you pay the model provider
- Platforms
- Linux, Docker
- Licence
- MIT
Research-first UX; not a daily-driver editor.
