RpCli

AI-powered coding CLI (rc) — interactive chat with local tools, plan/yolo modes, single prompts, and git commit messages.

Features

• 🤖 Interactive chat – natural language coding assistant with tool access
• 🔧 Built-in tools – file read/write/edit, shell commands, search, todos
• 📋 Todo management – track progress with todo_add, todo_split, todo_update, todo_list, todo_clear (see note under Interactive Chat)
• 🧠 Plan mode – read-only exploration before implementing changes
• 🔄 Model switching – switch between default and Expert models while carrying conversation context into a fresh session
• ⚡ YOLO mode – execute tools without confirmation
• 📝 Commit message generation – generate conventional commit messages from staged or all changes
• 🌐 HTTP server – serve an OpenAI-compatible API with rc serve

Install

npm install --global @rezaparsian/rp-cli
# or
pnpm add -g @rezaparsian/rp-cli

On first run, RP-CLI asks for a DeepSeek token and stores it in ~/.config/rp-cli/.env.

Usage

rc                            Open interactive chat
rc "your question"            Send a single prompt without thinking
rc -t "your question"         Enable and display thinking
rc -tq "your question"        Enable thinking but display only the final answer
rc -s "your question"         Enable web search
rc -st "your question"        Enable web search and thinking
rc -c                         Generate a commit message from staged changes
rc -c -a                      Generate a commit message from all changes (HEAD)
rc serve                      Start an OpenAI-compatible HTTP API
rc serve --port 8080          Listen on a custom port
rc --plan "your question"     Plan mode (read-only) for a single prompt

Options

| Option | Description | |--------------------------|--------------------------------------------------------| | -t, --thinking | Enable thinking for a single prompt. | | -q, --quiet | Hide thinking output. Combine with -t as -tq. | | -s, --search | Enable web search for a single prompt. | | -c, --commit-message | Generate a commit message from staged changes. | | -a, --commit-all | Use all changes from HEAD instead of staged changes. | | -p, --port | Port for rc serve (default: 3000). | | --host | Bind address for rc serve (default: 127.0.0.1). |

To save only the final answer while allowing the model to think:

rc -tq "write documentation for this project" > documentation.md

OpenAI-compatible API

rc serve exposes a local OpenAI Chat Completions API in front of DeepSeek. Point any OpenAI SDK or client at it.

rc serve
rc serve --port 8080 --host 127.0.0.1

| Endpoint | Description | |----------------------------------------|---------------------------------------| | POST /v1/chat/completions | Continue the current conversation. | | POST /v1/chat/completions/:sessionId | Continue a specific DeepSeek session. | | GET /health | Liveness check. |

Models: deepseek-chat (default) and deepseek-reasoner (thinking / reasoning_content). Requests use the token from ~/.config/rp-cli/.env, or Authorization: Bearer <token> if no config token is set.

curl http://127.0.0.1:3000/v1/chat/completions \
  -H "Content-Type: application/json" \
  -d '{"model":"deepseek-chat","messages":[{"role":"user","content":"Hello"}]}'

from openai import OpenAI

client = OpenAI(base_url="http://127.0.0.1:3000/v1", api_key="not-needed")
print(client.chat.completions.create(
    model="deepseek-chat",
    messages=[{"role": "user", "content": "Hello"}],
).choices[0].message.content)

The server keeps a persistent conversation and sends only the final item in each messages array. Use the :sessionId endpoint to select an existing DeepSeek conversation. stream: true returns SSE chunks ending with data: [DONE]. Successful completion responses include the active session in the X-RP-Session-Id header. Extra fields thinking_enabled and search_enabled are accepted.

The OpenAI-compatible reasoning_effort field is also accepted. none and minimal disable DeepSeek thinking; low, medium, high, xhigh, and max enable it. DeepSeek exposes thinking as a boolean, so enabled effort levels do not produce different reasoning depths. When both fields are present, thinking_enabled takes precedence.

Client-defined function tools are supported through the OpenAI tools and tool_choice fields. RP-CLI returns finish_reason: "tool_calls" and does not execute these functions. Execute the requested function in the client, then send its result as the final message with role: "tool" and the matching tool_call_id. Tool definitions are cached per session, so clients may omit unchanged tools on later requests.

Interactive Chat

rc with no arguments opens a persistent TUI. Ctrl+C clears non-empty input first; press it again on an empty input to exit. Esc stops an in-progress generation.

The system prompt is sent as soon as interactive mode opens. If the program closes before you send a message, the unused chat session is deleted automatically.

Type @ to mention a file or folder from the workspace. Type / to pick a slash command.

Execution modes

Press TAB to cycle normal → yolo → plan. The status bar shows the current mode. (Some workflows also switch modes via the /plan, /normal, and /yolo slash commands below.)

| Mode | Behavior | |----------|----------------------------------------------------------------------------------------------------------------------------------------------------------------| | normal | Mutating tools (write_file, edit_file, delete_file, run_command) ask for y/n first. | | yolo | Mutating tools run without confirmation, except elevated/sudo commands, which always ask. | | plan | Read-only. The model may inspect the repo, then you get Start this plan? y leaves plan, restores the previous mode, and executes. n stays in plan. |

Slash commands

| Command | Description | |---------------------|----------------------------------------------------------| | /init | Generate or update AGENTS.md for this repository. | | /continue | Keep going after the 10-round tool limit. | | /clear | Start a new conversation (clears todos too). | | /help | List commands. | | /search | Toggle web search. | | /model | Switch between the default and Expert models. The current conversation is summarized and carried into the new session. | | /thinking | Toggle thinking (on by default in chat). | | /logging (/log) | Toggle saving transcripts under ~/.config/rp-cli/logs. | | /todos | Show current todo list. | | /plan | Switch to plan mode. | | /normal | Switch to normal mode. | | /yolo | Switch to YOLO mode. | | /exit (/quit) | Close the application. |

Thinking and final responses are shown as separate messages. Web search and chat logging are off by default.

The input supports forward Delete, Ctrl+Backspace and Ctrl+Delete for word deletion, and Ctrl+Left/Right for word navigation. Alt+Backspace and Alt+Delete are disabled. Ctrl+J inserts a newline.

Todo tools

For multistep tasks, the agent can track progress with a built-in todo system:

1. Create a main task with todo_add
2. Split it into subtasks with todo_split (subtasks provided as a list, one per line)
3. Mark progress with todo_update as each subtask completes
4. Show the full list after every operation, or on demand with /todos

Example workflow:

User: "Build a REST API with Express and TypeScript"

Agent: *Creates todo #1 with 8 subtasks*
Agent: *Works through each subtask, updating status as it goes*
Agent: *Shows progress after each step*

Note: this todo system is documented in one version of the project's README but not the other — confirm it's still present before relying on it.

Single prompt

rc "..." sends one question, renders the response as formatted Markdown, then exits. Thinking is disabled by default. Use -t to enable and display it, or -tq to enable it silently and render only the final answer. Web search is also disabled by default and can be enabled with -s. Boolean short flags can be combined:

rc -st "find the latest Node.js release and explain the changes"
rc -stq "research this topic and save only the final answer" > answer.md

Thinking and final answers stream to the terminal as they are generated. Raw tool call markup is hidden during streaming and replaced by a concise activity message.

Commit message

rc -c reads your staged git diff, generates a Conventional Commit message, shows it, and asks for confirmation before committing. -a uses git diff HEAD instead of staged changes.

AI tools

RP-CLI tells the model about local tools and executes tool calls returned by the model:

• list_directory(path?) lists a directory.
• read_file(path) reads a UTF-8 file up to 100 KiB.
• write_file(path, content) creates or overwrites a file.
• edit_file(path, old_text, new_text) performs a unique exact replacement.
• delete_file(path) deletes a file after user confirmation.
• search_files(query, path?) searches files and returns up to 50 matching lines.
• run_command(command) runs a shell command. Include sudo anywhere in the command to open an OS authorization dialog (UAC on Windows).

When the user asks for a file, the model is instructed to write it with write_file rather than paste the contents into the chat.

For each tool call, RP-CLI keeps the assistant's explanatory text in the chat and replaces the raw <tool_calls> markup with a human-readable activity message. The assistant can chain tool calls; after 10 rounds it stops and you can type /continue.

All paths are restricted to the directory where RP-CLI was started. To add another tool, register it in source/tools/registry.ts; the system prompt is generated from the same registry. Tool calls use this protocol:

<tool_calls>
  <invoke name="read_file">
    <parameter name="path">package.json</parameter>
  </invoke>
</tool_calls>

Write literal < and > inside parameter values. Do not convert them to < or >. The only sequences that must not appear unescaped in a value are </parameter>, </invoke>, and </tool_calls>.

Configuration

The CLI stores configuration in ~/.config/rp-cli/:

• .env – API keys and preferences (including logging and model selection)

The selected model is stored as RP_CLI_MODEL=default or RP_CLI_MODEL=expert.

• logs/ – chat transcripts (if logging is enabled)

Development

# Clone the repository
git clone https://github.com/RezaParsian/RpCli.git
cd RpCli

# Install dependencies
pnpm install

# Build
pnpm build

# Development watch mode
pnpm dev

# Run tests
pnpm test

# Format code
pnpm format

License

MIT