An AI coding assistant that lives in your terminalterminal

Coder CLI is an open-source, terminal-first AI coding assistant built in Go. It pairs a conversational agent with real developer tools — the filesystem, the shell, Git, task tracking, and a live view of the model's reasoning — all inside a keyboard-driven terminal interface. And because it speaks the Ollama API, it runs against any Ollama-compatible model, from a laptop running locally to a hosted GPU endpoint in the cloud.

The Model Chooses What You See

Most agent harnesses render a fixed transcript and force the model to type into it. Here, the model gets a presentation surface and actively chooses what goes on it. Code being modified belongs on a canvas instead of being quoted into chat. A project plan belongs in a task pane. Tool traffic belongs in an activity log. The interface is model output, not a rigid frame the model is trapped inside.

Generative Interfaces

The goal is to move past generated text to interfaces composed dynamically from session state, then scale that same behavior to surfaces beyond the terminal. What starts in the terminal is a testbed for a broader thesis: models in charge of how data is structured, formatted, and presented in the moment.

Three Experimental Models

The harness speaks to any OpenAI-compatible model or provider with your own key, so you stay in control of where your code goes. On top of that, the gateway at tjcoder.com exposes three experimental research models:

• ergo-flash for fast, economical answers.
• ergo-pro for balanced everyday work.
• ergo-genius for the deepest reasoning.

They are research models: a vehicle for the model-directed-interface work above, and a path toward academic partnerships and funding.

The Tool Layer

• File manipulation: read_file, write_file, edit_file.
• Execution: run_command, run_test, background_job, git_status.
• Agent controls: manage_model, set_reminder, ui_control.
• Web and research: a real Chrome instance driven over CDP.

Headless mode (coder -p "...") exposes the same toolset to scripts and CI.

Seven Specialized Agents

Instead of one overstuffed persona, Coder CLI ships seven purpose-built agents, each with its own scoped toolset, system prompt, and default model. You swap between them with /agent — no restart, no context loss.

Any OpenAI or Ollama-Compatible Model

Coder CLI is provider-light by design. Bring your own key to any OpenAI-compatible provider, or speak the Ollama API directly for every model in that ecosystem — Llama, Gemma, Qwen, MiniMax, Mistral, and beyond — whether it runs on your own hardware or behind a remote endpoint.

• Bring your own key: point the CLI at any OpenAI-compatible provider with your own API key, so your code and data stay where you choose.
• Local: point it at localhost:11434 and code entirely offline against a model running on your own machine.
• Cloud: pass --host to a remote Ollama-compatible endpoint and leverage hosted GPU acceleration.
• Gemini: a native Google Gemini provider is included for teams already on that stack.

Small-Model Reliability

A subtle but important detail: Coder CLI was built with the assumption that the model might be small. Its tool runner supports native tool-calls and a fenced-JSON fallback parser, so even a compact local model that fumbles the function-calling grammar can still reliably drive the filesystem and shell. That fallback is what makes the “works with any model” claim hold up in practice, not just in theory.

Beyond the Terminal

The agent's reach extends past the shell. A browser bridge drives a real Chrome instance over the Chrome DevTools Protocol — navigate, click, type, screenshot, and manage cookies and tabs across isolated profiles. It's what powers the social-researcher agent's ability to research Reddit and LinkedIn without ever repeating an upvote or message. Headless mode (coder -p "…") makes the same engine scriptable for CI pipelines, and an MCP client lets the agent reach third-party servers for tools it doesn't natively ship.

Install It in One Line

Distribution is deliberately frictionless — an npm package downloads the right prebuilt binary on install, and a one-line shell installer covers everything else, with an option to set up Ollama and pull a default model in the same step.

# npm
npm install -g @tjcoder/cli
# macOS (Homebrew)
brew install tjcoder-labs/tap/coder
# Linux
curl -fsSL https://raw.githubusercontent.com/tjcoder-labs/cli/main/install.sh | zsh

About the Author

TJ Coder AI Labs builds modern, AI-powered developer tooling — from autonomous agents to terminal-native interfaces — out of Salt Lake City, Utah. Coder CLI is open source under the MIT license and available on npm, Homebrew, and GitHub.