ChatGPT CLI is a powerful, multi-provider command-line interface for working with modern LLMs. It supports OpenAI, Azure, Perplexity, LLaMA, and more, and includes streaming, interactive chat, prompt files, image/audio I/O, MCP tool calls, and an experimental agent mode for multi-step tasks with safety and budget controls.

• Features
  • Prompt Support
    • Using the prompt flag
    • Example
    • Explore More Prompts
  • Agent Mode (ReAct + Plan/Execute)
    • Quick Start
    • Workdir Safety
    • Budgets and Policy
    • Logs
  • MCP Support
    • Overview
    • Examples
    • Headers and Authentication
    • MCP Session Management
    • How MCP Results Are Used
• Installation
  • Using Homebrew (macOS)
  • Direct Download
    • Apple Silicon
    • macOS Intel chips
    • Linux (amd64)
    • Linux (arm64)
    • Linux (386)
    • FreeBSD (amd64)
    • FreeBSD (arm64)
    • Windows (amd64)
• Getting Started
• Configuration
  • General Configuration
  • LLM Specific Configuration
  • Agent Configuration
    • Default Policy
  • Custom Config and Data Directory
    • Example for Custom Directories
    • Variables for interactive mode
  • Switching Between Configurations with --target
  • Azure Configuration
  • Perplexity Configuration
  • 302 AI Configuration
  • Command-Line Autocompletion
    • Enabling Autocompletion
    • Persistent Autocompletion
• Markdown Rendering
• Development
  • Using the Makefile
  • Testing the CLI
• Reporting Issues and Contributing
• Uninstallation
• Useful Links
• Additional Resources

• Streaming mode: Real-time interaction with the GPT model.
• Query mode: Single input-output interactions with the GPT model.
• Interactive mode: The interactive mode allows for a more conversational experience with the model. Prints the token usage when combined with query mode.
• Thread-based context management: Enjoy seamless conversations with the GPT model with individualized context for each thread, much like your experience on the OpenAI website. Each unique thread has its own history, ensuring relevant and coherent responses across different chat instances.
• Sliding window history: To stay within token limits, the chat history automatically trims while still preserving the necessary context. The size of this window can be adjusted through the context-window setting.
• Custom context from any source: You can provide the GPT model with a custom context during conversation. This context can be piped in from any source, such as local files, standard input, or even another program. This flexibility allows the model to adapt to a wide range of conversational scenarios.
• Agent mode (ReAct + Plan/Execute): Run multi-step tasks that can think, act, and observe using tools like shell, file operations, and LLM reasoning. Supports both iterative ReAct loops and Plan/Execute workflows, with built-in * budget limits* (time, steps, tokens) and policy enforcement (allowed tools, denied commands, workdir sandboxing) for safe-by-default automation.
• Web search: Allow compatible models (e.g. gpt-5+) to fetch live web data during a query. Enable with the web setting and tune results using web_context_size.
• MCP (Model Context Protocol) support: Call external MCP tools via HTTP(S) or STDIO, inject their results into the conversation context, and continue the prompt seamlessly.
  • MCP session management: Built-in support for stateful MCP servers. The CLI automatically initializes sessions, attaches session identifiers, and renews them when they become invalid.
• Support for images: Upload an image or provide an image URL using the --image flag. Note that image support may not be available for all models. You can also pipe an image directly: pngpaste - | chatgpt "What is this photo?"
• Generate images: Use the --draw and --output flags to generate an image from a prompt (requires image-capable models like gpt-image-1).
• Edit images: Use the --draw flag with --image and --output to modify an existing image using a prompt ( e.g., "add sunglasses to the cat"). Supported formats: PNG, JPEG, and WebP.
• Audio support: You can upload audio files using the --audio flag to ask questions about spoken content. This feature is compatible only with audio-capable models like gpt-4o-audio-preview. Currently, only .mp3 and .wav formats are supported.
• Transcription support: You can also use the --transcribe flag to generate a transcript of the uploaded audio. This uses OpenAI’s transcription endpoint (compatible with models like gpt-4o-transcribe) and supports a wider range of formats, including .mp3, .mp4, .mpeg, .mpga, .m4a, .wav, and .webm.
• Text-to-speech support: Use the --speak and --output flags to convert text to speech (works with models like gpt-4o-mini-tts). If you have afplay installed (macOS), you can even chain playback like this:

  chatgpt --speak "convert this to audio" --output test.mp3 && afplay test.mp3

• Model listing: Access a list of available models using the -l or --list-models flag.
• Advanced configuration options: The CLI supports a layered configuration system where settings can be specified through default values, a config.yaml file, and environment variables. For quick adjustments, various --set-<value> flags are provided. To verify your current settings, use the --config or -c flag.

We’re excited to introduce support for prompt files with the --prompt flag in version 1.7.1! This feature allows you to provide a rich and detailed context for your conversations directly from a file.

The --prompt flag lets you specify a file containing the initial context or instructions for your ChatGPT conversation. This is especially useful when you have detailed instructions or context that you want to reuse across different conversations.

To use the --prompt flag, pass the path of your prompt file like this:

chatgpt --prompt path/to/your/prompt.md "Use a pipe or provide a query here"

The contents of prompt.md will be read and used as the initial context for the conversation, while the query you provide directly will serve as the specific question or task you want to address.

Here’s a fun example where you can use the output of a git diff command as a prompt:

git diff | chatgpt --prompt ../prompts/write_pull-request.md

In this example, the content from the write_pull-request.md prompt file is used to guide the model's response based on the diff data from git diff.

For a variety of ready-to-use prompts, check out this awesome prompts repository. These can serve as great starting points or inspiration for your own custom prompts!

ChatGPT CLI includes an experimental agent mode that can plan and run multi-step tasks using tools (shell, file ops, and LLM reasoning), while enforcing budget + policy constraints.

There are two agent modes:

• ReAct (-agent-mode react): iterative “think → act → observe” loop
• Plan/Execute (-agent-mode plan): generates a plan first, then executes it step-by-step

ReAct mode (default):

chatgpt why is my test failing? --agent

Plan/Execute mode:

chatgpt what is the weather like in brooklyn --agent --agent-mode plan

Agent file access can be restricted to a working directory. This is useful to prevent accidental reads/writes outside a project.

chatgpt "what files are in the /tmp directory" \
  --agent \
  --agent-work-dir .

If a step tries to read/write outside the workdir, it will be denied by policy (e.g. kind=path_escape).

Agent execution is governed by:

• Budget limits (iterations, steps, tool calls, wall-time, token usage)
• Policy rules (allowed tools, denied shell commands, file op allowlist, and workdir path restrictions)

This keeps the agent useful while still being safe-by-default.

When running in agent mode, ChatGPT CLI automatically writes detailed execution logs to the cache directory, under:

$OPENAI_CACHE_HOME/agent/

These logs include:

• Planner output (for Plan/Execute mode)
• Tool calls and their results
• Timing and budget usage
• Debug-level traces when debug logging is enabled

Each agent run gets its own timestamped log directory, making it easy to inspect what happened after the fact or debug unexpected behavior.

This is especially useful when:

• An agent run fails due to budget or policy limits
• You want to understand why the agent chose certain steps
• You’re developing or tuning agent policies and budgets

ChatGPT CLI supports the Model Context Protocol (MCP) over HTTP(S). This allows the CLI to call an MCP tool, inject the tool’s result into the current thread as context, and then run your prompt — all in one command. The integration is provider-agnostic.

You provide:

• MCP endpoint URL (--mcp)
• Tool name (--mcp-tool)
• Optional HTTP headers (--mcp-header)
• Tool arguments (--mcp-param or --mcp-params)

When --mcp is set, the CLI will:

1. POST a JSON-RPC tools/call request to your MCP server
2. Automatically initialize and manage an MCP session if required
3. Extract the tool output
4. Store it as an assistant message in the active thread (prefixed with [MCP: <tool>])
5. Submit your query to the model (if you provided one)

Local FastMCP echo server (minimal MCP HTTP example):

chatgpt \
  --mcp "http://127.0.0.1:8000/mcp" \
  --mcp-tool echo \
  --mcp-param 'payload={"foo":"bar","count":3,"enabled":true}' \
  "What did the MCP server receive?"

Apify MCP example (production MCP server):

chatgpt \
  --mcp "https://mcp.apify.com/?tools=epctex/weather-scraper" \
  --mcp-tool "epctex-slash-weather-scraper" \
  --mcp-header "Authorization: Bearer $APIFY_API_KEY" \
  --mcp-param locations='["Brooklyn, NY"]' \
  --mcp-param timeFrame=today \
  --mcp-param units=imperial \
  --mcp-param proxyConfiguration='{"useApifyProxy":true}' \
  --mcp-param maxItems=1 \
  "what should I wear today"

Using --mcp-params (raw JSON) instead of multiple --mcp-param flags:

chatgpt \
  --mcp "https://your-mcp-server.example.com" \
  --mcp-tool "some-tool-name" \
  --mcp-params '{"locations":["Brooklyn, NY"],"timeFrame":"today"}' \
  "what should I wear today"

Local MCP server over stdio (no HTTP, runs as a subprocess):

chatgpt \
  --mcp "stdio:python test/mcp/stdio/mcp_stdio_server.py" \
  --mcp-tool echo \
  --mcp-param 'payload={"foo":"bar","count":3}' \
  "What did the MCP server receive?"

MCP does not mandate a specific authentication mechanism. Some servers use Bearer tokens, others use API keys, cookies, or no auth at all. Use --mcp-header to pass whatever your MCP server requires:

--mcp-header "Authorization: Bearer $TOKEN"
--mcp-header "X-Api-Key: $API_KEY"

Some MCP servers require a session identifier (commonly mcp-session-id) to be established before tool calls are accepted. The ChatGPT CLI automatically manages MCP sessions for HTTP(S) servers that require them:

• Initializes a session when needed
• Caches the session identifier per endpoint
• Attaches it to subsequent requests
• Automatically re-initializes the session if the server invalidates it

You can explicitly pass a session header yourself using --mcp-header. If you do, the CLI will respect it and skip automatic session handling.

Tool results are injected into the conversation thread as context before your query runs. The injected message is stored as an assistant message and prefixed like this:

[MCP: <tool-name>] ...

If you run MCP without providing a query, the CLI will inject the context and exit:

chatgpt \
  --mcp "https://your-mcp-server.example.com" \
  --mcp-tool "some-tool-name" \
  --mcp-params '{"foo":"bar"}'

You can install chatgpt-cli using Homebrew:

brew tap kardolus/chatgpt-cli && brew install chatgpt-cli

For a quick and easy installation without compiling, you can directly download the pre-built binary for your operating system and architecture:

curl -L -o chatgpt https://github.com/kardolus/chatgpt-cli/releases/latest/download/chatgpt-darwin-arm64 && chmod +x chatgpt && sudo mv chatgpt /usr/local/bin/

curl -L -o chatgpt https://github.com/kardolus/chatgpt-cli/releases/latest/download/chatgpt-darwin-amd64 && chmod +x chatgpt && sudo mv chatgpt /usr/local/bin/

curl -L -o chatgpt https://github.com/kardolus/chatgpt-cli/releases/latest/download/chatgpt-linux-amd64 && chmod +x chatgpt && sudo mv chatgpt /usr/local/bin/

curl -L -o chatgpt https://github.com/kardolus/chatgpt-cli/releases/latest/download/chatgpt-linux-arm64 && chmod +x chatgpt && sudo mv chatgpt /usr/local/bin/

curl -L -o chatgpt https://github.com/kardolus/chatgpt-cli/releases/latest/download/chatgpt-linux-386 && chmod +x chatgpt && sudo mv chatgpt /usr/local/bin/

curl -L -o chatgpt https://github.com/kardolus/chatgpt-cli/releases/latest/download/chatgpt-freebsd-amd64 && chmod +x chatgpt && sudo mv chatgpt /usr/local/bin/

curl -L -o chatgpt https://github.com/kardolus/chatgpt-cli/releases/latest/download/chatgpt-freebsd-arm64 && chmod +x chatgpt && sudo mv chatgpt /usr/local/bin/

Download the binary from this link and add it to your PATH.

Choose the appropriate command for your system, which will download the binary, make it executable, and move it to your /usr/local/bin directory (or %PATH% on Windows) for easy access.

1. Set the OPENAI_API_KEY environment variable to your ChatGPT secret key. To set the environment variable, you can add the following line to your shell profile (e.g., ~/.bashrc, ~/.zshrc, or ~/.bash_profile), replacing your_api_key with your actual key:

   export OPENAI_API_KEY="your_api_key"

2. To enable history tracking across CLI calls, create a ~/.chatgpt-cli directory using the command:

   mkdir -p ~/.chatgpt-cli

   Once this directory is in place, the CLI automatically manages the message history for each "thread" you converse with. The history operates like a sliding window, maintaining context up to a configurable token maximum. This ensures a balance between maintaining conversation context and achieving optimal performance.

   By default, if a specific thread is not provided by the user, the CLI uses the default thread and stores the history at ~/.chatgpt-cli/history/default.json. You can find more details about how to configure the thread parameter in the Configuration section of this document.

3. Try it out:

   chatgpt what is the capital of the Netherlands

4. To start interactive mode, use the -i or --interactive flag:

   chatgpt --interactive

   If you want the CLI to automatically create a new thread for each session, ensure that the auto_create_new_thread configuration variable is set to true. This will create a unique thread identifier for each interactive session.

5. To use the pipe feature, create a text file containing some context. For example, create a file named context.txt with the following content:

   Kya is a playful dog who loves swimming and playing fetch.

   Then, use the pipe feature to provide this context to ChatGPT:

   cat context.txt | chatgpt "What kind of toy would Kya enjoy?"

6. To list all available models, use the -l or --list-models flag:

   chatgpt --list-models

7. For more options, see:

   chatgpt --help

The ChatGPT CLI adopts a four-tier configuration strategy, with different levels of precedence assigned to flags, environment variables, a config.yaml file, and default values, in that respective order:

1. Flags: Command-line flags have the highest precedence. Any value provided through a flag will override other configurations.
2. Environment Variables: If a setting is not specified by a flag, the corresponding environment variable (prefixed with the name field from the config) will be checked.
3. Config file (config.yaml): If neither a flag nor an environment variable is set, the value from the config.yaml file will be used.
4. Default Values: If no value is specified through flags, config.yaml, or environment variables, the CLI will fall back to its built-in default values.

You can also use flags, for example:

chatgpt "what files are here?" --agent --agent-work-dir /tmp

allowed_tools: [shell, llm, files]
denied_shell_commands: [rm, sudo, dd, mkfs, shutdown, reboot]
allowed_file_ops: [read, write]

By default, ChatGPT CLI stores configuration and history files in the ~/.chatgpt-cli directory. However, you can easily override these locations by setting environment variables, allowing you to store configuration and history in custom directories.

To change the default configuration or data directories, set the appropriate environment variables:

export OPENAI_CONFIG_HOME="/custom/config/path"
export OPENAI_DATA_HOME="/custom/data/path"
export OPENAI_CACHE_HOME="/custom/cache/path"

If these environment variables are not set, the application defaults to ~/.chatgpt-cli for configuration files and ~ /.chatgpt-cli/history for history.

You can maintain multiple configuration files side by side and switch between them using the --target flag. This is especially useful if you use multiple LLM providers (like OpenAI, Perplexity, Azure, etc.) or have different contexts or workflows that require distinct settings.

How it Works

When you use the --target flag, the CLI loads a config file named:

config.<target>.yaml

For example:

chatgpt --target perplexity --config

This will load:

~/.chatgpt-cli/config.perplexity.yaml

If the --target flag is not provided, the CLI falls back to:

~/.chatgpt-cli/config.yaml

Example Setup

You can maintain the following structure:

~/.chatgpt-cli/
├── config.yaml # Default (e.g., OpenAI)
├── config.perplexity.yaml # Perplexity setup
├── config.azure.yaml # Azure-specific config
└── config.llama.yaml # LLaMA setup

Then switch between them like so:

chatgpt --target azure "Explain Azure's GPT model differences"
chatgpt --target perplexity "What are some good restaurants in the Red Hook area"

Or just use the default:

chatgpt "What's the capital of Sweden?"

CLI and Environment Interaction

• The value of --target is never persisted — it must be explicitly passed for each run.
• The config file corresponding to the target is loaded before any environment variable overrides are applied.
• Environment variables still follow the name: field inside the loaded config, so name: perplexity enables PERPLEXITY_API_KEY.

• %date: The current date in the format YYYY-MM-DD.
• %time: The current time in the format HH:MM:SS.
• %datetime: The current date and time in the format YYYY-MM-DD HH:MM:SS.
• %counter: The total number of queries in the current session.
• %usage: The usage in total tokens used (only works in query mode).

The defaults can be overridden by providing your own values in the user configuration file. The structure of this file mirrors that of the default configuration. For instance, to override the model and max_tokens parameters, your file might look like this:

model: gpt-3.5-turbo-16k
max_tokens: 4096

This alters the model to gpt-3.5-turbo-16k and adjusts max_tokens to 4096. All other options, such as url , completions_path, and models_path, can similarly be modified.

You can also add custom HTTP headers to all API requests. This is useful when working with proxies, API gateways, or services that require additional headers:

custom_headers:
  X-Custom-Header: "custom-value"
  X-API-Version: "v2"
  X-Client-ID: "my-client-id"

If the user configuration file cannot be accessed or is missing, the application will resort to the default configuration.

Another way to adjust values without manually editing the configuration file is by using environment variables. The name attribute forms the prefix for these variables. As an example, the model can be modified using the OPENAI_MODEL environment variable. Similarly, to disable history during the execution of a command, use:

OPENAI_OMIT_HISTORY=true chatgpt what is the capital of Denmark?

This approach is especially beneficial for temporary changes or for testing varying configurations.

Moreover, you can use the --config or -c flag to view the present configuration. This handy feature allows users to swiftly verify their current settings without the need to manually inspect the configuration files.

chatgpt --config

Executing this command will display the active configuration, including any overrides instituted by environment variables or the user configuration file.

To facilitate convenient adjustments, the ChatGPT CLI provides flags for swiftly modifying the model, thread , context-window and max_tokens parameters in your user configured config.yaml. These flags are --set-model , --set-thread, --set-context-window and --set-max-tokens.

For instance, to update the model, use the following command:

chatgpt --set-model gpt-3.5-turbo-16k

This feature allows for rapid changes to key configuration parameters, optimizing your experience with the ChatGPT CLI.

For Azure, you need to configure these, or similar, value

name: azure
api_key: <your azure api key>
url: https://<your_resource>.openai.azure.com
completions_path: /openai/deployments/<your_deployment>/chat/completions?api-version=<your_api>
auth_header: api-key
auth_token_prefix: " "

You can set the API key either in the config.yaml file as shown above or export it as an environment variable:

export AZURE_API_KEY=<your_key>

For Perplexity, you will need something equivelent to the following values:

name: perplexity
api_key: <your perplexity api key>
model: sonar
url: https://api.perplexity.ai

You can set the API key either in the config.yaml file as shown above or export it as an environment variable:

export PERPLEXITY_API_KEY=<your_key>

You can set the API key either in the config.yaml file as shown above or export it as an environment variable:

export AZURE_API_KEY=<your_key>

I successfully tested 302.AI with the following values

name: ai302 # environment variables cannot start with numbers
api_key: <your 302.AI api key>
url: https://api.302.ai

You can set the API key either in the config.yaml file as shown above or export it as an environment variable:

export AI302_API_KEY=<your_key>

Enhance your CLI experience with our new autocompletion feature for command flags!

Autocompletion is currently supported for the following shells: Bash, Zsh, Fish, and PowerShell. To activate flag completion in your current shell session, execute the appropriate command based on your shell:

• Bash

  . <(chatgpt --set-completions bash)

• Zsh

  . <(chatgpt --set-completions zsh)

• Fish

  chatgpt --set-completions fish | source

• PowerShell

  chatgpt --set-completions powershell | Out-String | Invoke-Expression

For added convenience, you can make autocompletion persist across all new shell sessions by adding the appropriate sourcing command to your shell's startup file. Here are the files typically used for each shell:

• Bash: Add to .bashrc or .bash_profile
• Zsh: Add to .zshrc
• Fish: Add to config.fish
• PowerShell: Add to your PowerShell profile script

For example, for Bash, you would add the following line to your .bashrc file:

. <(chatgpt --set-completions bash)

This ensures that command flag autocompletion is enabled automatically every time you open a new terminal window.

You can render markdown in real-time using the mdrender.sh script, located here. You'll first need to install glow.

Example:

chatgpt write a hello world program in Java | ./scripts/mdrender.sh

To start developing, set the OPENAI_API_KEY environment variable to your ChatGPT secret key.

The Makefile simplifies development tasks by providing several targets for testing, building, and deployment.

• all-tests: Run all tests, including linting, formatting, and go mod tidy.

  make all-tests

• binaries: Build binaries for multiple platforms.

  make binaries

• shipit: Run the release process, create binaries, and generate release notes.

  make shipit

• updatedeps: Update dependencies and commit any changes.

  make updatedeps

For more available commands, use:

make help

.\scripts\install.ps1

1. After a successful build, test the application with the following command:

   ./bin/chatgpt what type of dog is a Jack Russel?

2. As mentioned previously, the ChatGPT CLI supports tracking conversation history across CLI calls. This feature creates a seamless and conversational experience with the GPT model, as the history is utilized as context in subsequent interactions.

   To enable this feature, you need to create a ~/.chatgpt-cli directory using the command:

   mkdir -p ~/.chatgpt-cli

If you encounter any issues or have suggestions for improvements, please submit an issue on GitHub. We appreciate your feedback and contributions to help make this project better.

If for any reason you wish to uninstall the ChatGPT CLI application from your system, you can do so by following these steps:

If you installed the CLI using Homebrew you can do:

brew uninstall chatgpt-cli

And to remove the tap:

brew untap kardolus/chatgpt-cli

If you installed the binary directly, follow these steps:

1. Remove the binary:

   sudo rm /usr/local/bin/chatgpt

2. Optionally, if you wish to remove the history tracking directory, you can also delete the ~/.chatgpt-cli directory:

   rm -rf ~/.chatgpt-cli

1. Navigate to the location of the chatgpt binary in your system, which should be in your PATH.

2. Delete the chatgpt binary.

3. Optionally, if you wish to remove the history tracking, navigate to the ~/.chatgpt-cli directory (where ~ refers to your user's home directory) and delete it.

Please note that the history tracking directory ~/.chatgpt-cli only contains conversation history and no personal data. If you have any concerns about this, please feel free to delete this directory during uninstallation.

• Amazing Prompts
• OpenAI API Reference
• OpenAI Key Usage Dashboard
• OpenAI Pricing Page
• Perplexity API Reference
• Perplexity Key Usage Dashboard
• Perplexity Models
• 302.AI API Reference

• "Summarize any text instantly with a single shortcut" on Medium: Dive deep into the capabilities of this CLI tool with this detailed walkthrough.
• Join the conversation on Reddit: Discuss the tool, ask questions, and share your experiences with our growing community.

Thank you for using ChatGPT CLI!