A split image: hands weaving glowing threads on one side, a small conductor before a bright tower on the other, suggesting building and directing.
An open-source coding agent is a tool you both build with and direct: you keep the code and the model choice, and the agent does the weaving.

The proprietary coding tools get the headlines, but a large open-source field now does the same job: Claude Code , Cursor , and OpenAI Codex all have free, source-available alternatives you can inspect, self-host, and point at any model. This page maps that field. It covers the terminal agents, the editor extensions, the autonomous platforms, and the persistent orchestrators, and it explains which one fits which job.

Every tool here runs the same core loop: read the code, plan a change, edit files, run commands, and check the result. That shared shape is why they feel similar. What actually separates them are three axes, and if you hold these in mind the whole field sorts itself out.

  • Surface. Where you sit: a pure terminal interface, an extension inside your IDE, or a server you reach from a chat app.
  • Model coupling. Whether the tool is locked to one vendor’s models (like Codex or Gemini CLI) or is provider-agnostic and runs against Claude, GPT, Gemini, or a local model (like OpenCode, Aider, or Cline).
  • Autonomy. Whether it pauses for your approval on every action (Cline) or runs a task end to end in a sandbox (OpenHands).

The shared anatomy

Under the surface differences, these agents are built from the same layers.

Interface
Terminal (TUI) IDE extension Chat platforms Where you give instructions and review changes
Agent loop
Plan Edit Run Check The read-plan-edit-run cycle over your repository
Tools and permissions
File edits Shell Browser MCP servers Approval gates
Models
Anthropic, OpenAI, Google Open-weight models Local runtimes Provider-agnostic tools swap this layer freely

The field at a glance

MaintainerLicenseModelsSurfaceBest for
OpenCodeAnomaly (SST)MITAny providerTerminal, IDE, webProvider choice, multi-surface
AiderAider communityApache-2.0Any providerTerminalGit-centric pair coding
GooseAAIF (Linux Foundation)Apache-2.0Any providerDesktop, CLI, APIOn-machine automation
CrushCharmFSL-1.1-MITAny providerTerminalOne fast Go binary
ClineCline BotApache-2.0Any providerVS Code, CLIReviewed edits in an IDE
ContinuecontinuedevApache-2.0Any providerIDE, CLIAutocomplete plus agent
OpenHandsAll-Hands-AIMIT (mixed)Any via LiteLLMCLI, cloud, CIAutonomous sandboxed tasks
Codex CLIOpenAIApache-2.0OpenAI modelsTerminalOpenAI-native terminal work
Gemini CLIGoogleApache-2.0Gemini modelsTerminalSearch-grounded, big context
Qwen CodeAlibaba (QwenLM)Apache-2.0Any providerTerminalOpen Qwen-model coding
Hermes AgentNous ResearchMITAny providerCLI, 20+ chat appsPersistent orchestrator
Roo CodeRoo Code, Inc.Apache-2.0Any providerVS CodeDiscontinued 2026-05-15

The terminal-native agents

These run in your shell and edit your local files. OpenCode is provider-agnostic and unusual for its client/server split, with a headless server that a terminal, desktop, IDE, or web client can all attach to, plus a build agent and a read-only plan agent. Aider is the git-native option: it commits every change as a separate, reversible git commit and builds a map of your repository to work on larger codebases. Goose is a broader on-machine agent with a desktop app, CLI, and API, extended through MCP and packaged into shareable YAML recipes.

Crush is worth knowing because of a naming tangle. The original OpenCode was a Go project (opencode-ai/opencode); it was archived, and its author continued the work at Charm under the name Crush. So Crush is the official continuation of that original codebase: a single Go binary with strong LSP-enhanced context and mid-session model switching. The OpenCode covered on this wiki is a separate TypeScript agent that carries the opencode name today.

Three of the terminal agents come from the big model labs and ship as open-source CLIs, but two are tied to their maker’s models. Codex CLI (OpenAI, Apache-2.0, written in Rust) runs against OpenAI models via a ChatGPT account or an API key. Gemini CLI (Google, Apache-2.0) is locked to Gemini models but adds Google Search grounding and a very large context window. Qwen Code (Alibaba, Apache-2.0) is the exception: it began as a fork of Gemini CLI (from version 0.8.2, then developed independently) and stays provider-agnostic, working with OpenAI, Anthropic, Gemini, and Qwen models as well as local runtimes.

The IDE and platform agents

Cline lives mainly in VS Code and is built around a human in the loop: it plans in Plan mode, executes in Act mode, and asks approval for every file edit and terminal command, with checkpoints to undo its work. Continue is the long-standing open-source assistant across VS Code and JetBrains, offering autocomplete, chat, and an agent, plus a command-line agent called cn. Note one caveat: its main continuedev/continue repository is now read-only and marked no longer actively maintained, while its command-line agent ships from the separate @continuedev/cli package. Check the current state before you standardise a team on it.

OpenHands sits at the autonomous end. Formerly OpenDevin, it runs the agent inside a sandboxed Docker container so it can write code, run commands, and browse safely, and it offers a CLI, a cloud service, and a GitHub Action for issue-to-pull-request automation.

Roo Code was a popular fork of Cline that added a multi-mode system (Code, Architect, Ask, Debug, and custom modes). Its official documentation states the extension was shut down on 2026-05-15 and points users to Cline or a community fork. Treat it as discontinued rather than a current choice.

The orchestrator: a different layer

Hermes Agent from Nous Research does not belong in the same column as the others. It is a persistent, memory-backed agent that runs as a server and reaches you across more than 20 messaging platforms, and it does deep coding by delegating to OpenCode or Codex rather than editing files itself. If the tools above are the workers, Hermes is closer to a foreman that remembers context and hands out the work. See multi-agent orchestration for the pattern it uses.

How to choose

Ask 1 Where do you work? Terminal points to OpenCode or Aider; an IDE points to Cline.
Ask 2 Which models? Locked to one vendor is fine for Codex or Gemini CLI; otherwise pick provider-agnostic.
Ask 3 How much autonomy? Approve each step with Cline, or run tasks end to end with OpenHands.
Your priorityBest fit
Provider freedom in the terminalOpenCode, Aider
Reviewed edits inside an IDECline
Autonomous, sandboxed runsOpenHands
On-machine general automationGoose
A single fast Go binaryCrush
An always-on assistant with memoryHermes Agent
Staying in one vendor’s ecosystemCodex CLI, Gemini CLI

The open-source tools trade the polish and support of a paid product for two things that matter to many teams: you can point them at any model, including a local one, and you can read and self-host the code. If those matter less to you than a single integrated experience, the proprietary trio may still be the better pick. Many teams run both: a paid tool for daily editing and an open-source agent for provider-flexible or self-hosted work.

Further reading

Sources