Coding agent (CLI) · he-yufeng
CoreCoder review
Quick verdict. CoreCoder: Deliberately small Python coding agent: a 1.5k-line engine in about 3.3k readable lines. Suits developers who want to learn how a coding agent works by reading a small, working one, or who want a base to fork. A weaker fit for teams who need a polished, supported coding assistant with a large feature set.
CoreCoder review
CoreCoder is a coding agent made small on purpose. Its author describes it as the nanoGPT of coding agents: an engine of about 1,515 lines inside 3,343 lines of pure Python, meant to be read in an afternoon so that you can see how a tool like Claude Code or Aider works underneath, and then fork it. This review follows the GitHub README and repository, read on 18 November 2026, and we did not run an independent test. It has no separate vendor site, so the repository is the official source.
What it does
It runs a normal agent loop in the terminal: corecoder opens an interactive session, and corecoder -p "task" runs one task and exits. It has eight tools (bash, read, write and edit file, glob, grep, a todo list and sub-agent spawning), plan mode that refuses every change until you approve a plan, hooks around tool calls, checkpoints and three-tier context compaction. Anything that would change your disk or run a command stops for your consent first, and one-shot mode refuses those tools unless you pass --yes.
It speaks the OpenAI-compatible API by default, so switching to DeepSeek, a local Ollama model or another provider is usually two environment variables. MCP servers, local or remote over Streamable HTTP, can be added from ~/.corecoder/mcp.json. The repository also holds eight source-reading essays and a suite of offline tests.
Price and licence
CoreCoder is free under the MIT licence; you pay only your model provider. It installs with pip install corecoder and needs Python 3.10 or newer. The repository showed 1,924 stars on 18 November 2026 and was created on 1 April 2026.
Limits
The author says plainly that the point is to learn from it and fork it, not to make it your daily driver, and the README lists gaps on purpose: there is no retrieval index for large repositories. There is no graphical app or hosted plan. It is a single-author project, and a high star count shows interest, not production use. The README mentions Git Bash routing for Windows but gives no platform matrix, so check your setup before relying on it. If you want a finished tool, see the open-source CLI coding agents list.
Facts: platforms, price and licence
- platforms
- macOSLinuxWindowsPure Python 3.10+ command-line tool (`pip install corecoder`). The repo has shell routing for POSIX shells, with Git Bash on Windows. No separate website was found; the GitHub project is the official source.
- price
- FreeFree and open source (MIT). You pay your model provider; it speaks the OpenAI-compatible API by default and can point at DeepSeek, a local Ollama or other endpoints through environment variables (README).
- source
- Open sourceLicence: MIT
- github
- he-yufeng/CoreCoder 1.9k stars, seen 2026-11-22
CoreCoder features
- ✓ yes MCP
Not verified: Multi-agent, Headless / CI, Sandbox, Local models, Your own API key, Git worktrees, SSH. How to read feature data.
CoreCoder strengths
- The whole engine (loop, model interface, context, tools, sessions) is about 1,515 lines inside 3,343 lines of Python, small enough to read in an afternoon (README)
- It really runs: eight tools (bash, read/write/edit file, glob, grep, a todo list and sub-agent spawning), plan mode, hooks, checkpoints and three-tier context compaction
- Anything that changes your disk or runs a command stops for your consent first; `-p` one-shot mode refuses those tools unless you pass `--yes`
- Works with any OpenAI-compatible endpoint, and MCP servers (local, or remote over Streamable HTTP) can be added from `~/.corecoder/mcp.json`
- Eight source-reading essays in the repo explain how each part works
CoreCoder weak points
- The author says it is meant to be read and forked, not to become your daily driver; it is not a replacement for a mature coding agent
- It keeps only a minimal core: no retrieval index for large repositories (the README lists this as open)
- No separate vendor site, GUI or hosted service; you bring the model and key
- A single-author project (he-yufeng); the star count is not evidence of production use
Who CoreCoder is for
- best for
- Developers who want to learn how a coding agent works by reading a small, working one, or who want a base to fork.
- look elsewhere if
- Teams who need a polished, supported coding assistant with a large feature set.
Lists that include CoreCoder
Similar tools
Sources
Checked 2026-10-21. Prices and features change; the vendor's page is the authority.
- github.com/he-yufeng/CoreCoder — README: size, tools, providers, MCP, plan mode, limits; licence, stars and dates from the GitHub API
Not verified: Hands-on behaviour not tested; Windows support rests on the repo's Git Bash routing; the README does not give a platform matrix