ccr

Free terminal coding assistant. Reads your repo, proposes diffs, and runs shell commands with your approval. No API key required.

Install

npm install -g @ryanisavibecoder/ccr

Quick Start

ccr login                         # sign up — opens your browser
ccr "explain this codebase"       # one-shot
ccr                               # interactive REPL

That's it. ccr login creates a free account and provisions everything. The service routes your requests across multiple LLM providers (Groq, Together AI, Cerebras, OpenRouter) so individual users don't need to manage API keys.

Free tier

Every account gets 2,000 requests per month. Usage is shown inline in the REPL header and at the end of every one-shot run. Resets on the 1st of each month UTC.

Auth options

ccr login                # browser flow (default)
ccr login --terminal     # email + password directly in the terminal
ccr login --method github # force GitHub OAuth (browser only)

Credentials are stored at ~/.ccr/auth.json (mode 0600).

Configure

# Override the default model:
export CCR_MODEL=llama-3.3-70b-versatile

# Point at a self-hosted or staging deployment:
export CCR_ENDPOINT=https://my-ccr.example.com

Bring your own key (advanced)

If you'd rather skip the managed service and use Groq directly:

export GROQ_API_KEY=gsk_...
ccr "your prompt"

This bypasses the proxy. You'll get a deprecation warning on each run but it still works for offline / power-user setups.

Use

ccr                                  # interactive Ink UI in current dir
ccr "explain this project"           # interactive UI, prompt pre-submitted
ccr -p "list TODOs"                  # one-shot console mode (no UI)
ccr --resume                         # resume most-recent session
ccr --list-sessions                  # show saved sessions
ccr --yolo                           # auto-approve writes/shell
ccr --model llama-3.3-70b-versatile  # override model
ccr --cwd /path/to/repo "..."        # work in another dir

Interactive UI features:

• Persistent input prompt at the bottom; messages stream above.
• Spinner while the model is thinking; Ctrl-C to interrupt the run, press it again to exit.
• Tool calls render as inline cards: ◌ tool_name(args…) → ✓ or ✗.
• Approval prompts appear inline as a yellow box with a colorized diff; press y to approve, n or Esc to deny.
• Slash commands: /help, /clear, /sessions, /save, /exit.

Tools the model can call

Read-only (auto-allowed):

• read_file(path, offset?, limit?)
• glob(pattern, path?)
• grep(pattern, path?, glob?)

Modifying (each shows a diff/command and asks for approval):

• write_file(path, content)
• edit_file(path, old_string, new_string, replace_all?)
• multi_edit(path, edits[]) — atomic sequence of search-replace edits
• insert_lines(path, line, content) — insert at a 1-indexed line
• bash(command, timeout?)

Project context

These files at the project root are loaded as system context: CLAUDE.md, AGENTS.md, .ccr/context.md.

Sessions

Persisted to ~/.ccr/sessions/<project-id>/<timestamp>.json. Resume with ccr --resume (latest) or ccr --resume <id>.

Limitations

• No long-running watcher loop yet.
• No memory-distillation layer; transcripts are kept verbatim per session.
• Single-project sessions; no cross-project recall.