🦾 Coding Agent Harness

Before installingCopy Anchor Link

Run the installation commands in VS Code's integrated terminal, not in Claude Code's chat input.

The shell examples below target macOS, Linux, or Windows through WSL. Use the linked project documentation for native Windows installation.

• Install Node.js and npm for the npm-based commands.
• Pi requires Node.js 22.19 or newer.
• Install Python and pip for Aider and mini-swe-agent.
• Run each agent from the project directory where you want it to work.
• Configure model authentication separately for each agent.
• Review remote installer scripts before executing commands that pipe downloaded content directly into a shell.

Check the tools available in your terminal:

node --version
npm --version
python --version
python -m pip --version

If your system uses python3 instead of python, substitute python3 in the Python commands below.

DeepAgents CodeCopy Anchor Link

DeepAgents Code, launched with dcode, is a terminal coding agent built on the Deep Agents SDK.

CLI installationCopy Anchor Link

Install using the official installer:

curl -LsSf [https://langch.in/dcode](https://langch.in/dcode) | bash

Start it from your project directory:

dcode

Configure credentials for your chosen model provider during setup.

Installing the deepagents Python library alone is not the same as installing the dcode terminal application.

Best featuresCopy Anchor Link

• Persistent memory carries project conventions and learned patterns across conversations.
• Task-specific subagents support parallel delegation.
• Remote sandboxes can execute tools away from your local machine.
• Goals and rubrics provide explicit criteria for checking completion.
• Context compaction summarizes older messages and offloads originals to storage.
• LangSmith tracing helps inspect and debug agent operations.

PiCopy Anchor Link

Pi is a minimal, extensible agent harness designed to adapt to your workflow.

CLI installationCopy Anchor Link

Install globally with npm:

npm install -g --ignore-scripts @earendil-works/pi-coding-agent

Start Pi from your project directory:

pi

Inside Pi, connect a supported subscription or API key:

/login

Alternatively, use the official macOS/Linux installer, which can install the required Node.js version if needed:

curl -fsSL [https://pi.dev/install.sh](https://pi.dev/install.sh) | sh

Choose either the npm method or the installer rather than running both.

Best featuresCopy Anchor Link

• Extensions, skills, prompt templates, and themes customize behavior and presentation.
• Pi packages bundle customizations for distribution through npm or Git.
• A TypeScript SDK lets you embed the agent in your own application.
• RPC supports external programmatic control.
• Print and JSON modes support automation outside an interactive terminal.
• The monorepo includes reusable model, agent-runtime, and terminal-UI libraries.

OpenCodeCopy Anchor Link

OpenCode provides an interactive coding workflow with configurable primary agents and subagents.

CLI installationCopy Anchor Link

Install the current OpenCode CLI package:

npm install -g @opencode/cli@latest

Verify the installation and launch it:

opencode --version
opencode

Alternatively, use the official OpenCode 2 installer:

curl -fsSL [https://opencode.ai/v2/install](https://opencode.ai/v2/install) | bash

Choose one installation method.

The older opencode-ai package and the current package both use the opencode command. Avoid mixing installations that compete for the same executable.

Best featuresCopy Anchor Link

• Build and Plan agents provide distinct implementation and analysis workflows.
• Specialized subagents support repository exploration and delegated work.
• Custom agents can use different models, prompts, and permissions.
• JSON and Markdown definitions make agent configuration easy to version with a project.
• Child-session navigation lets you inspect delegated work separately from the main conversation.
• Permission policies control which actions run automatically, require approval, or are denied.

Codex CLICopy Anchor Link

Codex CLI is OpenAI’s open-source coding agent that runs locally on your computer.

CLI installationCopy Anchor Link

Install globally with npm:

npm install -g @openai/codex

Start Codex from your project directory:

codex

Follow the sign-in flow to connect your ChatGPT account, or configure API-key authentication using the official documentation.

Alternatively, use the official macOS/Linux standalone installer:

curl -fsSL [https://chatgpt.com/codex/install.sh](https://chatgpt.com/codex/install.sh) | sh

Choose one installation method.

Best featuresCopy Anchor Link

• A local terminal interface supports repository-based coding workflows.
• Sandbox and approval controls distinguish permitted execution from actions requiring additional authorization.
• Editor integration options support workflows in VS Code and related editors.
• The app-server interface provides a programmatic foundation for clients and integrations.

Gemini CLICopy Anchor Link

Gemini CLI is an open-source terminal agent with code, multimodal, and automation capabilities.

CLI installationCopy Anchor Link

Install the stable release globally with npm:

npm install -g @google/gemini-cli@latest

Start Gemini CLI from your project directory:

gemini

Alternatively, try it without a persistent global installation:

npx @google/gemini-cli

Follow the authentication instructions for your account.

The repository currently positions this CLI around enterprise Code Assist and legacy paid API or Vertex access. It directs free individual and Google AI Pro/Ultra users toward Antigravity CLI. Check current account eligibility before configuring access.

Best featuresCopy Anchor Link

• Multimodal context can include images, PDFs, and sketches.
• MCP servers extend the agent with external tools.
• Non-interactive execution supports scripts and automated workflows.
• JSON and streaming JSON output support machine-readable integrations.
• GEMINI.md context files provide persistent project instructions.
• Documentation covers checkpointing, sandboxing, trusted folders, and a VS Code companion.

AiderCopy Anchor Link

Aider focuses on AI pair programming inside an existing Git repository.

CLI installationCopy Anchor Link

Install and run Aider's installation helper:

python -m pip install aider-install
aider-install

Start Aider from your project directory:

aider

Configure your selected model and provider credentials before making model requests. Model-specific setup is documented in Aider's model connection guide.

Best featuresCopy Anchor Link

• A repository map supplies structural context about the codebase.
• Automatic Git commits make generated changes easier to inspect and undo.
• Automatic linting and testing can feed failures back into the editing workflow.
• Support for cloud and local models gives you deployment flexibility.
• Images and web pages can provide additional context.
• Voice input supports spoken coding requests.

mini-swe-agentCopy Anchor Link

mini-swe-agent is a deliberately small software-engineering agent designed to be understandable and easy to modify.

CLI installationCopy Anchor Link

Create an isolated Python environment and install the package:

python -m venv ~/.venvs/mini-swe-agent
source ~/.venvs/mini-swe-agent/bin/activate
python -m pip install mini-swe-agent

Start the CLI from your project directory:

mini

In a new terminal, reactivate the environment before running it:

source ~/.venvs/mini-swe-agent/bin/activate
mini

Alternatively, try it in a temporary managed environment using uv:

python -m pip install uv
uvx mini-swe-agent

Configure the model and provider credentials using the project's quickstart instructions.

Best featuresCopy Anchor Link

• The agent class is roughly 100 lines of Python; the complete project also includes model, environment, and execution support.
• Bash is the sole action interface, avoiding a large catalog of specialized tools.
• Linear message history makes execution trajectories easier to inspect.
• Independent command execution simplifies swapping execution backends.
• Local and containerized environments support different testing and isolation needs.
• A CLI, Python bindings, batch inference, and a trajectory browser support experimentation.

Which should you try first?Copy Anchor Link

These are workflow recommendations, not benchmark rankings.

Using these alongside Claude Code in VS CodeCopy Anchor Link

These tools are separate agents, not Claude Code plugins.

• Run interactive agents in separate VS Code terminal tabs.
• Test one agent at a time before introducing concurrent workflows.
• Use separate branches or worktrees when multiple agents may edit files.
• Configure authentication independently for each tool.
• Start with repository analysis before allowing edits or shell execution.
• Use external isolation when tool permissions alone do not provide the boundaries you need.

How to compare them fairlyCopy Anchor Link

Use a fixed task set rather than GitHub stars as your quality measure.

1. Use the same repository revision and task description.
2. Keep the model and budget consistent where possible.
3. Record task completion, test results, elapsed time, and usage cost.
4. Track human interventions and permission approvals.
5. Inspect the resulting diff for unrelated changes.
6. Repeat tasks before drawing conclusions from a single run.

When tools require different models, label the result as a comparison of complete agent systems rather than harnesses alone.