MetaClaw Evolving Agent

Skill by ara.so — Daily 2026 Skills collection

MetaClaw is an OpenAI-compatible proxy agent that intercepts conversations, injects learned skills, and continuously improves itself through real-world interactions. It supports three modes: lightweight skills injection, immediate RL training, and a smart "madmax" scheduler that defers weight updates to idle/sleep windows.

Installation

# Minimal — skills injection only, no GPU required
pip install -e .

# Full RL training support (torch, transformers, tinker)
pip install -e ".[rl]"

# Skill evolution via LLM summarization
pip install -e ".[evolve]"

# Google Calendar scheduler for madmax mode
pip install -e ".[scheduler]"

# Recommended: everything
pip install -e ".[rl,evolve,scheduler]"

Quick Start

# One-time interactive config wizard
metaclaw setup

# Start in default madmax mode (skills + RL + smart scheduler)
metaclaw start

# Skills only — no GPU, no Tinker needed
metaclaw start --mode skills_only

# RL mode — trains immediately when batch is full
metaclaw start --mode rl

# RL without scheduler (same as above, explicit)
metaclaw start --mode rl

After metaclaw start, a local OpenAI-compatible proxy is running. Point your client (OpenClaw or any OpenAI SDK consumer) at http://localhost:<port> instead of the upstream LLM endpoint.

Configuration

metaclaw setup writes a config file (default: ~/.metaclaw/config.yaml). You can also edit it directly:

# ~/.metaclaw/config.yaml

proxy:
  host: 0.0.0.0
  port: 8080

llm:
  provider: kimi          # kimi | qwen | claude | minimax | openai | gemini
  base_url: https://api.moonshot.cn/v1
  model: moonshot-v1-8k
  # api_key loaded from env: METACLAW_LLM_API_KEY

skills:
  enabled: true
  max_injected: 5         # max skills injected per turn
  summarize_after_session: true

rl:
  enabled: true
  backend: auto           # auto | tinker | mint
  batch_size: 32
  algorithm: grpo
  opd_teacher: false      # optional teacher distillation

scheduler:                # madmax mode only
  enabled: true
  sleep_hours: [22, 7]    # local 22:00–07:00
  idle_timeout_minutes: 15
  google_calendar: false  # set true + configure OAuth for meeting detection

logging:
  level: info
  log_dir: ~/.metaclaw/logs

Environment Variables

export METACLAW_LLM_API_KEY="your-llm-api-key"
export METACLAW_TINKER_API_KEY="your-tinker-api-key"   # rl mode
export METACLAW_MINT_API_KEY="your-mint-api-key"        # if backend=mint
export GOOGLE_CALENDAR_CREDENTIALS_PATH="path/to/creds.json"  # scheduler

Operating Modes

Python API

Programmatic startup

import asyncio
from metaclaw import MetaClawAgent, AgentConfig, Mode

async def main():
    config = AgentConfig.from_yaml("~/.metaclaw/config.yaml")
    agent = MetaClawAgent(config, mode=Mode.MADMAX)
    await agent.start()

asyncio.run(main())

Manual skill injection

from metaclaw.skills import SkillStore, SkillInjector

store = SkillStore(path="~/.metaclaw/skills")

# Add a skill manually
store.add(
    name="code-review-checklist",
    content="Always check for: 1) error handling, 2) type hints, 3) docstrings.",
    tags=["code", "review"]
)

# Retrieve top-k relevant skills for a query
injector = SkillInjector(store)
relevant = injector.retrieve(query="review my Python function", top_k=3)
for skill in relevant:
    print(skill.name, skill.score)

Intercepting and recording conversations

from metaclaw.proxy import ConversationInterceptor
from metaclaw.memory import ExperienceBuffer

buffer = ExperienceBuffer(max_size=1000)

interceptor = ConversationInterceptor(
    upstream_url="https://api.moonshot.cn/v1",
    on_complete=buffer.record   # called after each turn with (messages, response)
)

# buffer.record signature:
async def on_complete(messages: list[dict], response: dict) -> None:
    ...

Triggering RL training manually

from metaclaw.training import RLTrainer, TrainingConfig

trainer = RLTrainer(
    config=TrainingConfig(
        backend="tinker",       # or "mint"
        algorithm="grpo",
        batch_size=32,
        lora_rank=16,
    )
)

# Collect a batch from the experience buffer and train
async def run_training(buffer):
    batch = buffer.sample(n=32, split="support")   # support/query separation
    result = await trainer.train(batch)
    print(f"Training complete. Loss: {result.loss:.4f}, Steps: {result.steps}")

Reward modeling

from metaclaw.rewards import RewardModel

reward_model = RewardModel(provider="llm")  # uses configured LLM for scoring

async def score_turn(prompt: str, response: str) -> float:
    score = await reward_model.score(prompt=prompt, response=response)
    return score  # float in [-1.0, 1.0]

Skills Lifecycle

Conversation turn
       │
       ▼
 SkillInjector.retrieve()   ← vector search over SkillStore
       │  injects top-k skills into system prompt
       ▼
 LLM responds
       │
       ▼
 ExperienceBuffer.record()  ← stores (context, response, metadata)
       │
       ▼ (end of session)
 SkillSummarizer.run()      ← LLM extracts reusable patterns
<