Minimal and Customizable CC-Style Coding Agent. That's the entire thing.

A fully-functional CC-style AI coding assistant you can read in an afternoon — and tell it to extend itself.
"Oh My Code!!!"

Why OhMyCode?

🤏 It's tiny. OhMyCode delivers the CC-style coding assistant experience — streaming, tool use, context compression, memory — in ~3000 lines you fully own.

🧬 OhMyCode extends itself. Tell OhMyCode to "add a tool that does X" and it edits this repo's source to build one — no plugin SDK, no API boundaries between you and the code.

Highlights

• 🧩 Minimal CC Core — Streaming output, tool execution, permission modes, context compression, memory, and resume in ~3000 lines.
• 🔧 Deeply Customizable — Add your own tools/providers/prompts or just ask OhMyCode to extend itself by editing its own source.
• 🎒 Skills Included — Comes with practical built-in skills for add-tool, add-provider, add-feature, debugging, workflow conventions, test generation, and benchmarking.
• 🌐 Provider Flexibility — Works with OpenAI, Anthropic, Azure, and OpenAI-compatible APIs.

Install

git clone <repo-url>
cd OhMyCode
./scripts/setup-cli.sh

Tip

This project is CLI-first. Start it with ohmycode command only. If command is still not found after setup, reload shell and verify: source ~/.zshrc && which ohmycode

./scripts/setup-cli.sh installs a stable global shim at ~/.local/bin/ohmycode, so it works consistently across different shell sessions and conda environments.

Quick Start

ohmycode                        # interactive REPL
ohmycode -p "Fix the bug"      # single-shot prompt
ohmycode --resume               # resume last conversation
ohmycode --mode plan            # read-only mode

Verify CLI wiring (recommended for agents and CI scripts):

pip3 show ohmycode | rg "Editable project location|Location"
which ohmycode
ohmycode --help

REPL Commands

Command	Description
/exit	Quit (auto-saves + extracts memories)
/clear	Clear history
/mode <mode>	Switch mode (default / auto / plan)
/memory list|delete	Manage memories
/skills	List skills
/<skill-name>	Run a skill

Configuration

Create ~/.ohmycode/config.json:

{
  "provider": "openai",
  "model": "gpt-4o",
  "api_key": "sk-...",
  "mode": "auto"
}

Config merges four layers: system defaults < user (~/.ohmycode/) < project (.ohmycode/) < CLI args.

CLI overrides: --provider, --model, --mode, --api-key, --base-url.

All config options

Key	Default	Description
provider	"openai"	LLM provider
model	"gpt-4o"	Model name
mode	"default"	Permission mode
base_url	""	API base URL
api_key	""	API key
azure_endpoint	""	Azure endpoint
azure_api_version	"2024-02-01"	Azure API version
max_turns	100	Max conversation turns
token_budget	200000	Token budget
output_tokens_reserved	8192	Reserved output tokens
rules	[]	Permission rules
system_prompt_append	""	Appended to system prompt
search_api / search_api_key	""	Web search API config

Built-in Tools

Tool	Description
bash	Shell commands with timeout
read	Read files with line range
edit	Find-and-replace (unique match)
write	Create/overwrite files
glob	Find files by pattern
grep	Regex search across files
web_fetch	Fetch URL → text
web_search	DuckDuckGo search
agent	Sub-agent (max depth 2)

Extending (or: Let It Extend Itself)

The fastest way to add a feature? Ask OhMyCode to do it. It can read its own source, write new files, and run tests — so "add a tool that counts words" is a single prompt away.

You can also do it manually:

• Add a tool — create ohmycode/tools/my_tool.py, auto-discovered. See /add-tool skill.
• Add a provider — create ohmycode/providers/my_provider.py, auto-discovered. See /add-provider skill.
• Add a skill — create SKILL.md in .ohmycode/skills/, .claude/skills/, .agents/skills/, or ~/.ohmycode/skills/ (searched in that order).

Project Structure

ohmycode/
├── cli.py               # REPL + rendering
├── core/
│   ├── loop.py          # Conversation loop
│   ├── messages.py      # Message types + streaming events
│   ├── context.py       # Token counting + compression
│   ├── permissions.py   # Permission pipeline
│   └── system_prompt.py # System prompt assembly
├── providers/           # OpenAI, Anthropic (auto-discovered)
├── tools/               # 9 built-in tools (auto-discovered)
├── skills/loader.py     # Skill scanner
├── memory/memory.py     # B+-Tree memory + LLM extraction
├── storage/conversation.py  # JSON persistence + resume
└── config/config.py     # Four-layer config merge
benchmarks/
├── run_bench.py         # Benchmark harness (token tracking + scoring)
└── suite.py             # 8 SWE-bench-style task definitions

Testing

python3 -m pytest tests/ -v          # 190 unit tests

Benchmarking

OhMyCode ships with a built-in benchmark suite — 8 SWE-bench-style coding tasks that test code generation, bug fixing, refactoring, test generation, tool use, and code comprehension.

python3 benchmarks/run_bench.py                        # full benchmark, current config
python3 benchmarks/run_bench.py --model gpt-4o-mini    # compare a different model
python3 benchmarks/run_bench.py --dry-run              # validate tasks without LLM

Or from the REPL: /bench

The harness tracks token usage (in/out) per task and outputs a scorecard + bench_results.json for model comparison.

Development Closed-Loop

OhMyCode includes skills for a TDD-style development loop:

code  →  /gen-tests  →  /run-tests  →  fix  →  repeat

Skill	Purpose
/gen-tests <module>	Generate tests following project conventions
/run-tests [scope]	Run tests, analyze failures, fix, re-run
/bench	Score any provider/model with 8 agent tasks

License

MIT