PipeAgent: A command-line llm agent tool over pipe

Most agent can be rebuilt with pipe in Unix Philosophy: data streaming in and out. We revert complex tools into the most simple forms: API (function) call, and build this tool.

Table of Contents

• pipe-agent
  • Table of Contents
  • Features
  • Plans
  • Quick Startup
  • Installation
    • From PyPI
    • From Source Code
  • Configuration
    • Overriding Configurations
  • Usage
    • Basic Query
    • Prompt Concatenation
      • Using -- to separate arguments
    • Using a different model
    • Using Context Chaining
  • Running Tests

Features

• Unix-like CLI: Rich and useful command line interface.

• Prompt Concatenation: Concatenate multiple prompt input sources via command-line arguments, files, or standard input.

• Context Chaining: Maintain conversation context across multiple calls and connect with pipe.

• Chain-of-Thought (CoT): Automatically filter out reasoning content from the final output.

• Streaming (SSE): Get real-time responses from the API.

• Flexible Configuration: Support multiple configuration file input with overriding.

• Shell Auto-completion: Integrated support for argcomplete for easy use.

• Agent over shell: Build your own agent with pipe-agent and text processing utilities and tricks.

Plans

• Refine models and provider logic.
• Refine documentation.
• Support Multimodal LLMs.
• Interface Change Refine concatenation logic.
• Features Support workflow-based agent through configuration files.

Quick Startup

1. pip install .
2. pipe-agent and then configures the model in ~/.config/pipe-agent/models.yaml and provider configs in ~/.config/pipe-agent/models/.
3. pipe-agent hello world

Installation

From PyPI [Currently not Available]

python -m pip install --upgrade pipe-agent
# Recommend `pipx` for global installation 
pipe-agent Hello World

From Source Code

1. Clone the repository and install the package:

   git clone https://github.com/CreeperLKF/pipe-agent.git
   cd pipe-agent
   pip install .
   

2. Activate command-line completion: (Optional)

   Add the following to your .bashrc or .zshrc file for permanent auto-completion.

   eval "$(register-python-argcomplete pipe-agent)"
   

   You may need to restart your shell for the changes to take effect.

Configuration

On the first run, pipe-agent will automatically create a configuration directory at ~/.config/pipe-agent/ with default template files. You can override this path by setting the PIPE_AGENT_CONFIG_PATH environment variable.

The new configuration is split into multiple files:

• default.conf: For general, non-model related settings.
• models.yaml: Orchestrates which models and providers are active.
• models/: A directory containing individual YAML configuration files for each API provider (e.g., OpenAI.yaml).

models.yaml

This file controls which providers are enabled and defines metrics for model selection.

• providers.enabled: A list of provider names to activate. The order determines the priority when selecting a model without specifying a provider. For example, if both OpenAI and DeepSeek are listed and both provide a gpt-4 model, pipe-agent will use the one from the provider listed first.
• models.metrics: Defines different ways to rank models. The default metric is used when no model is specified. You can create custom metrics (e.g., cost, speed) and rank models accordingly.

Provider Configuration (e.g., models/OpenAI.yaml)

Each file in the models/ directory defines the settings for a specific provider.

• provider.api_base_url: The API endpoint for the provider.
• provider.openai_api_key: Configure one or more API keys, with support for sequential or random rotation and retry limits.
• provider.defaults: Set default parameters (like temperature or enable_sse) for all models from this provider.
• models: A list of models available from this provider, including their model_id, a user-friendly model_name, and optional model_alias list for easier command-line access.

Overriding Configurations

You can override settings from default.conf by using the -f flag. The tool loads configurations in this order:

1. default.conf (or default.local if exists)
2. Each file specified with -f path in the order they appear.

When a relative path is given to -f, the tool searches in this order:

1. ./path
2. ./path.local
3. $PIPE_AGENT_CONFIG_PATH/path
4. $PIPE_AGENT_CONFIG_PATH/path.local

The -f flag can be used multiple times. Settings in later files will override those in earlier ones.

Usage

This document is currently under construction. Please refer to pipe-agent --help for more information.

Basic Query

$ pipe-agent "What is love"
$ pipe-agent -p "prompt.txt"
$ pipe-agent "劲发江潮落，气收秋毫平！" -p "重岳1.txt" "千招百式在一息！" -p "重岳2.txt"

Basically, pipe-agent get prompts from -p (files) and positional arguments. Multiple file inputs and positional arguments inputs are supported (use -p path multiple times). If PROMPT_CONCAT_NL is true, a newline will be appended after prompt from -p. If PROMPT_CONCAT_SP is true, a space will be appended after prompt from a positional arguments.

You can also enter interactive mode by running pipe-agent without any prompt arguments. By default, you press Ctrl+D to send. Use the -n flag to send the prompt on the first Enter press instead.

$ pipe-agent -n
What is love?
<Assistant's response will appear here>

Or use pipe bump input into stdin.

$ echo "Summarize this text." | pipe-agent

If a prompt is provided via command-line arguments or a file, the content from stdin will be appended to the existing prompt.

$ echo "and this part from stdin" | pipe-agent "This is a prompt from arguments"
# Final prompt will be: "This is a prompt from arguments\nand this part from stdin"

Prompt Concatenation

Besides specifying multiple inputs, you can add text before or after your main prompt using the -B (before) and -A (after) flags. These can be used multiple times.

# Joins positional arguments with a space by default
$ pipe-agent I think Python "is a great language" -B "At any time, C++ is the best." -B "In my opinion," -A "What is the best language?" -A "(Only answer C++/Python according to the context)"
# Final Prompt: "At any time, C++ is the best.\nIn my opinion,\nI think Python is a great language\nWhat is the best language?\n(Only answer C++/Python according to the context)\n"

You can also set default PROMPT_BEFORE and PROMPT_AFTER values in your .conf file. The final prompt is assembled in this order: Config Before -> CLI Before -> Main Prompt -> CLI After -> Config After.

By default, parts are joined with a newline (controlled by PROMPT_CONCAT_NL=true|false). The positional arguments that form the Main Prompt are joined by a space (controlled by PROMPT_CONCAT_SP=true|false).

Using -- to separate arguments

You can use -- to tell pipe-agent to treat everything that follows as part of the prompt, even if it looks like a flag. It is recommended to use -- in automated jobs (agent or your shell scripts) for security concerns.

$ pipe-agent -m gpt4 -- -f is not a file, it is part of the prompt.

Using a different model

You can select a model using the -m or -M flags, which support tab-completion.

• -m model[@provider]: Selects a model primarily by its name or alias. If the model name is ambiguous (e.g., available from multiple providers), you must specify the provider.
• -M provider[@model]: Selects a model primarily by its provider.

# Select gpt-4 from the OpenAI provider
$ pipe-agent -m gpt-4@OpenAI "Tell me a joke."

# If 'gpt4t' is an alias for gpt-4-turbo, this also works
$ pipe-agent -m gpt4t "Summarize this article."

# Select a model from the DeepSeek provider
$ pipe-agent -M DeepSeek "What is your name?"

If no model is specified via the command line, the tool selects the highest-ranking model from the default metric in models.yaml.

Using Context Chaining

Start a conversation. The output will be only the assistant's response.

pipe-agent "My name is Bob. Remember it."

Continue the conversation by piping the history back in and getting only the new response. The -c o flag prints the full history and -c i flag accepts a history json from stdin (which means stdin cannot receive prompts in this mode).

# First turn, save full history to chat.json
pipe-agent "My name is Bob" -c o > chat.json

# Second turn, provide history via stdin, provide new prompt, get full history back
cat chat.json | pipe-agent "What is my name?" -c io > chat_updated.json

# You can also get just the final response
cat chat_updated.json | pipe-agent "Thank you." -c i

Running Tests

To run the test suite, first install the test dependencies:

pip install .[test]

Then, you can run the tests using pytest:

pytest

You can also generate a code coverage report:

pytest --cov=pipe_agent