clawagents 4.0.0

Python port of the ClawAgents framework

ClawAgents Python

A lean, full-stack agentic protocol. ~2,500 LOC.

Quick Start

pip install -e .

Create a .env:

PROVIDER=gemini
GEMINI_API_KEY=AIza...
GEMINI_MODEL=gemini-3-flash-preview
STREAMING=1
CONTEXT_WINDOW=128000
MAX_TOKENS=4096

One-Line Agent

from clawagents import create_claw_agent

agent = create_claw_agent("gemini-3-flash")
result = await agent.invoke("List all Python files in src/")
print(result.result)

With Instruction

agent = create_claw_agent("gpt-5", instruction="You are a code reviewer.")
result = await agent.invoke("Review the code and suggest improvements")

CLI

python -m clawagents --task "Find all TODO comments in the codebase"

API

create_claw_agent(model, instruction, ...)

Param	Type	Default	Description
model	str | LLMProvider | None	None	Model name or provider. None = auto-detect from env
instruction	str	None	What the agent should do / how it should behave
tools	list	None	Additional tools. Built-in tools always included
skills	str | list	auto-discover	Skill directories. Default: checks ./skills, ./.skills, ./skill, ./.skill
memory	str | list	auto-discover	Memory files. Default: checks ./AGENTS.md, ./CLAWAGENTS.md
streaming	bool	True	Enable streaming
on_event	callable	None	Event callback

Built-in Tools

Every agent includes these — no setup needed:

Tool	Description
ls	List directory with size + modified time
read_file	Read file with line numbers + pagination
write_file	Write/create file (auto-creates dirs)
edit_file	Replace text (supports replace_all)
grep	Search — single file or recursive with glob filter
glob	Find files by pattern (**/*.py)
execute	Shell command execution
write_todos	Plan tasks as a checklist
update_todo	Mark plan items complete
task	Delegate to a sub-agent with isolated context
use_skill	Load a skill's instructions (when skills exist)

Hooks (Convenience Methods)

agent = create_claw_agent("gemini-3-flash", instruction="Code reviewer")

# Block dangerous tools
agent.block_tools("execute", "write_file")

# Or whitelist only safe tools
agent.allow_only_tools("read_file", "ls", "grep", "glob")

# Inject context into every LLM call
agent.inject_context("Always respond in Spanish")

# Limit tool output size
agent.truncate_output(3000)

Advanced: Raw hooks are also available for custom logic:

agent.before_llm = lambda messages: messages      # modify messages before LLM
agent.before_tool = lambda name, args: True        # return False to block
agent.after_tool = lambda name, args, result: result  # modify tool results

Auto-Discovery

The factory automatically discovers project files:

What	Default locations checked
Memory	./AGENTS.md, ./CLAWAGENTS.md
Skills	./skills, ./.skills, ./skill, ./.skill, ./Skills

Pass explicit paths to override: memory="./docs/AGENTS.md", skills=["./my-skills", "./shared-skills"]

Memory System

Project Memory

Loads AGENTS.md files and injects content into every LLM call. Use for project context.

Auto-Compaction

When conversation exceeds 75% of CONTEXT_WINDOW:

1. Full history offloaded to .clawagents/history/compacted_*.json
2. Older messages summarized into [Compacted History]
3. Last 6 messages kept intact

Environment Variables

Variable	Default	Description
PROVIDER	auto-detect	openai or gemini
OPENAI_API_KEY	—	OpenAI API key
OPENAI_MODEL	gpt-5-nano	OpenAI model
GEMINI_API_KEY	—	Gemini API key
GEMINI_MODEL	gemini-3-flash-preview	Gemini model
STREAMING	1	1 = enabled, 0 = disabled
CONTEXT_WINDOW	128000	Token budget for compaction
MAX_TOKENS	4096	Max output tokens per response

Testing

pip install -e ".[dev]"
python -m pytest tests/ -v