coding-agent

A local, terminal-based coding agent — an LLM that reads, searches, and edits files in your repo and runs shell commands through an interactive approval gate. It runs entirely on your machine and talks to any OpenAI-compatible endpoint using your own API key (BYOK — bring your own key). Your code and your key never touch anyone else's servers; requests go only to the provider you configure.

Install

npm install -g @studiofrontier/coding-agent

Or run without installing:

npx @studiofrontier/coding-agent setup

Requires Node.js >= 20.

Configure (BYOK)

Run the one-time setup to store your provider credentials locally:

coding-agent setup

It prompts for three things:

• Base URL — any OpenAI-compatible endpoint, e.g. https://api.openai.com/v1, an OpenRouter URL, a Gemini OpenAI-compatible URL, or your own gateway.
• API key — your provider key (input is hidden; stored locally only).
• Model id — e.g. gpt-4o, gemini-2.0-flash, or whatever your endpoint exposes.

Credentials are written to ~/.config/coding-agent/config.json (or $XDG_CONFIG_HOME/coding-agent/config.json) with owner-only permissions (0600). The API key is sent only to the endpoint you configure.

Non-interactive / CI:

coding-agent setup --base-url https://api.openai.com/v1 --api-key sk-... --model gpt-4o
coding-agent setup --show    # print current config (key redacted)
coding-agent setup --path    # print the config file location

Configuration precedence

Each field is resolved from the first source that provides it:

| Priority | Source | Notes | |---|---|---| | 1 | Environment variables | CODING_AGENT_BASE_URL, CODING_AGENT_API_KEY, CODING_AGENT_MODEL (the LITELLM_* names also work as aliases) | | 2 | A .env in the current directory | Handy for per-project overrides | | 3 | The stored config | Written by coding-agent setup |

This means a project .env or an exported env var can override your global key, and CI can inject everything via environment variables.

Usage

Interactive REPL:

coding-agent

Slash commands inside the REPL: /help, /clear, /reset, /sessions [all], /resume <id>, /exit. While the agent is working you'll see a live spinner (⠋ Thinking… (3s), ⠋ Running bash… (1s)).

Multi-line input: end a line with a backslash (\) to continue onto a new line; a plain Enter submits. (Terminals don't send a distinct code for Shift+Enter, so this shell-style continuation is the portable way to write multi-line prompts.)

One-shot (pipeable — the answer goes to stdout, notices to stderr):

coding-agent -p "summarize what this project does"
coding-agent "add a CHANGELOG.md with an Unreleased section"

Options

  -p, --print          one-shot mode; prompt is read from the arguments
  -c, --continue       resume the most recent session for the current directory
      --resume <id>    resume a specific session by id (prefix ok)
      --model <id>     override the model for this run
      --cwd <dir>      workspace root (default: current directory)
      --max-steps <n>  max tool round-trips per turn (default: 25)
      --yolo           auto-approve all tool calls (no prompts)
      --verbose        debug logging to stderr
  -h, --help           show help

Sessions & history

• Input recall: press ↑/↓ in the REPL to recall prompts from previous sessions (persisted to ~/.config/coding-agent/history).
• Session resume: interactive REPL conversations are saved per project under ~/.config/coding-agent/sessions/. Resume the most recent one for the current directory with coding-agent -c, or a specific one with coding-agent --resume <id> (list them with /sessions). One-shot -p runs are ephemeral and are never saved. Resuming keeps a fresh system prompt for the current directory and replays the saved messages after it.

Safety

• File tools are confined to the workspace root — the agent cannot read or write outside the directory it's launched in (or --cwd).
• Shell commands are not sandboxed. Their safety boundary is the approval prompt: write_file, edit_file, and bash ask for confirmation (y / n / always-allow) before running. --yolo disables prompting — use it only in environments you trust.

License

ISC © Chetan Bhogade