# Code Puppy

> An open-source, MIT-licensed AI code agent for the terminal that generates, edits, and explains code with zero telemetry and support for 65+ LLM providers.

Code Puppy is an open-source, MIT-licensed AI-powered code generation agent built by Mike Pfaffenberger and installable via `uvx code-puppy`. It was created, per the README, "angrily in reaction to Windsurf and Cursor removing access to models and raising prices," and runs entirely from the command line with a strong privacy-first architecture. The project reports 826 GitHub stars and 170k+ PyPI downloads as of the available data.

## What It Is

Code Puppy is a terminal-based agentic coding assistant in the same category as Cursor and Windsurf, but designed to run without a proprietary IDE. It understands programming tasks, generates and edits code, executes shell commands, and explains its reasoning — all from a CLI prompt. Because it is 100% open source under the MIT license, users can inspect, fork, and self-host it without any vendor lock-in.

## Model Access and Provider Support

One of Code Puppy's headline features is its breadth of model support. It integrates with [models.dev](https://models.dev) to give access to 65+ providers and over 1,000 model offerings via a single `/add_model` interactive TUI command. Out of the box it supports OpenAI, Anthropic, Google Gemini, Cerebras, Ollama, xAI (Grok), Groq, Mistral, Together AI, Perplexity, DeepInfra, Cohere, and many more. A round-robin model distribution feature lets users cycle across multiple API keys or models to stay within rate limits. For full privacy, users can point it at a local VLLM, SGLang, or Llama.cpp server so no data leaves their network.

## Speculative Tool Execution

A notable performance feature called Speculative Tool Execution (toggled with `Ctrl+X Ctrl+S`) starts running the model's tool calls while the model is still generating them. Read-heavy operations like `read_file`, `grep`, and `list_files` launch the moment their argument line closes, hiding tool latency behind the model's own typing time. A live scoreboard above the prompt tracks hits, misses, and estimated time saved per session.

## Agent System and Customization

Code Puppy ships with a flexible multi-agent system. Users can switch between built-in Python agents (like the default `code-puppy` agent and an `agent-creator` helper) or define their own agents as JSON files stored in `~/.code_puppy/agents/`. JSON agents support custom system prompts, tool access lists, pinned models, and per-agent model settings — no Python knowledge required. The system also supports `AGENTS.md` files for project-level coding standards and conventions, compatible with the [agent.md](https://agent.md) specification.

## Privacy Architecture

The README describes a zero-telemetry commitment: no usage analytics, no prompt logging, no behavioral profiling, and no third-party data sharing. The project is independently maintained with no corporate or investor backing, which the author argues makes the privacy commitment structurally enforceable. Prompts flow only to the LLM provider the user configures, and a fully local inference path is supported.

## Platform and Setup

Code Puppy requires Python 3.11+ and is installable on macOS, Linux, Windows, and Android (via Termux). The recommended install method is `uvx code-puppy`, which requires no separate pip install step. An optional DBOS durable-execution plugin (installed via the `durable` extra) checkpoints agent runs to a local SQLite or Postgres database, enabling crash recovery and resumption of long sessions. MCP server integration is available via the `/mcp` command.

## Features
- AI-powered code generation and editing
- Terminal/CLI-based interface
- Speculative tool execution for reduced latency
- Support for 65+ LLM providers via models.dev integration
- Round-robin model distribution for rate limit management
- Multi-agent system with Python and JSON agent types
- Custom agent creation via JSON configuration
- AGENTS.md support for project coding standards
- MCP server integration
- Shell command execution
- File read/write/delete/grep tools
- DBOS durable execution for crash recovery
- Zero telemetry and privacy-first architecture
- Local inference support (VLLM, SGLang, Llama.cpp)
- Meta Muse OAuth integration
- Custom slash commands via markdown files
- Auto-compact with system message merging
- Context truncation via /truncate command

## Integrations
OpenAI, Anthropic, Google Gemini, Cerebras, Ollama, xAI (Grok), Groq, Mistral, Together AI, Perplexity, DeepInfra, Cohere, Meta Muse, models.dev, DBOS, MCP servers, Pydantic AI, VLLM, SGLang, Llama.cpp, OpenRouter, AIHubMix

## Platforms
WINDOWS, MACOS, LINUX, ANDROID, WEB, API, CLI

## Pricing
Open Source

## Version
pr-530-assets

## Links
- Website: https://github.com/mpfaffenberger/code_puppy
- Documentation: https://code-puppy.dev
- Repository: https://github.com/mpfaffenberger/code_puppy
- EveryDev.ai: https://www.everydev.ai/tools/code-puppy