micro-cc 0.1.1

Harness that gives frontier models full system access — shell, filesystem, browser, MCP — running directly on the metal.

micro·cc — cognitive compute

  ███╗   ███╗ ██████╗ ██████╗ ███████╗██╗
  ████╗ ████║██╔═══██╗██╔══██╗██╔════╝██║
  ██╔████╔██║██║   ██║██║  ██║█████╗  ██║
  ██║╚██╔╝██║██║   ██║██║  ██║██╔══╝  ██║
  ██║ ╚═╝ ██║╚██████╔╝██████╔╝███████╗███████╗
  ╚═╝     ╚═╝ ╚═════╝ ╚═════╝ ╚══════╝╚══════╝
  ─────── on the metal ─────────────────────

Harness that gives frontier models full system access — shell, filesystem, browser, MCP — running directly on the metal. Planning, memory, and tool orchestration let models perform complex multi-step work autonomously. Point it at any project directory and go.

Install

pip install micro-cc

Environment

Create ~/.micro-cc/.env with one endpoint — set Anthropic OR LiteLLM, not both. If both are set, Anthropic takes priority.

# Anthropic (direct)
ANTHROPIC_API_KEY=sk-ant-

# OR LiteLLM (proxy to Bedrock, Azure, etc.)
LITELLM_BASE_URL=https://your-proxy.com
LITELLM_API_KEY=sk-...

# Shared
OPENAI_API_KEY=sk-proj-
SERPAPI_KEY=

Usage

microcc /path/to/your/project

Controls:

• Enter — submit prompt
• Shift+Enter — newline
• Escape — interrupt
• /clear — reset conversation
• /model — switch model
• /exit — quit

Data

All conversation history and project state is stored in:

~/.micro-cc/
  .env                              API keys (you create this)
  projects/
    {project}_{hash}/
      messages.jsonl                conversation history
      summary.json                  sliding-window summary
      project_path.txt              maps hash back to directory

Architecture

start_live_.py (CLI)                textual TUI + rich rendering
       │
       │ async for event in claude_loop()
       ▼
claude_loop_.py (Core)              API calls, tool execution, JSONL storage
       │
       │ execute_tool_call()
       ▼
tools/                              bash_, read_, write_, edit_, glob_, grep_
       │
       ▼
~/.micro-cc/projects/{hash}/        conversation persistence

How it works:

1. CLI starts FileWatcher on project dir, enters prompt loop
2. User query → claude_loop() async generator
3. Builds system prompt with project's CLAUDE.md + any file changes detected
4. Calls Anthropic API with tool definitions
5. Executes tool calls locally, loops until the model stops calling tools
6. Yields events (thinking, tool_call, tool_result, final_text) for CLI to render
7. Saves conversation to JSONL (stripped of thinking blocks for replay)

Tools:

Tool	Description
bash_	Subprocess execution in project_dir, timeout support
read_	Read files with line numbers, offset/limit
write_	Create/overwrite files, auto-creates dirs
edit_	Surgical string replacement (fails if ambiguous)
glob_	Find files by pattern, sorted by mtime
grep_	Regex search with context lines

Path handling: Relative paths resolve to project_dir, absolute paths work anywhere.

File watcher: Detects external changes, injects into system prompt so the model sees what you edited.