Inspired by how the brain learns. Meta-learn and evolve your 🦞 from every conversation in the wild. No GPU required. Works with Kimi, Qwen, Claude, MiniMax, and more.
⚡ Supported LLM Providers & Platforms
Kimi |
Qwen |
Claude |
MiniMax |
OpenAI |
Gemini |
+ Much More |
🧬 RL Training Backends for Weight Evolution
Tinker |
MinT |
More Coming |
🇨🇳 中文 • 🇯🇵 日本語 • 🇰🇷 한국어 • 🇫🇷 Français • 🇩🇪 Deutsch • 🇪🇸 Español • 🇧🇷 Português • 🇷🇺 Русский • 🇮🇹 Italiano • 🇻🇳 Tiếng Việt • 🇦🇪 العربية • 🇮🇳 हिन्दी
Overview • Quick Start • Configuration • Skills Mode • RL Mode • MadMax Mode • Citation
metaclaw setup # one-time config wizard
metaclaw start # default: madmax mode — skills + scheduled RL training
metaclaw start --mode rl # RL without scheduler (trains immediately on full batch)
metaclaw start --mode skills_only # skills only, no RL (no Tinker needed)
- [03/13/2026] v0.3.1 — MinT backend support: RL training now works with both Tinker and MinT. Configurable via
rl.backend(auto/tinker/mint). - [03/13/2026] v0.3 — Continual meta-learning support: slow RL updates now only run during sleep hours, idle time, or Google Calendar meetings. Added support/query set separation to prevent stale reward signals from polluting model updates.
- [03/11/2026] v0.2 — One-click deployment via
metaclawCLI. Skills enabled by default, RL is now opt-in. - [03/09/2026] We release MetaClaw — Just talk to your agent and let it evolve automatically. NO GPU deployment required; just plug into the API.
video_v2_compressed.mp4
MetaClaw is an agent that meta-learns and evolves in the wild. Just talk to your agent as you normally would — MetaClaw turns every live conversation into a learning signal, enabling the agent to continuously improve through real-world deployment rather than offline training alone.
Under the hood, it places your model behind an OpenAI-compatible proxy that intercepts interactions from OpenClaw, injects relevant skills at each turn, and meta-learns from accumulated experience. Skills are summarized automatically after each session; with RL enabled, a meta-learning scheduler defers weight updates to idle windows so the agent is never interrupted during active use.
No GPU cluster required. MetaClaw works with any OpenAI-compatible LLM API out of the box, and uses a Tinker-compatible backend for cloud-based LoRA training. Tinker is the default reference path, and MinT can be enabled through a separate compatibility package when needed.
Configure once with metaclaw setup, then metaclaw start brings up the proxy, injects skills, and wires OpenClaw automatically. No manual shell scripts needed.
| Mode | Default | What it does |
|---|---|---|
skills_only |
Proxy your LLM API. Skills injected and auto-summarized after each session. No GPU/Tinker required. | |
rl |
Skills + RL training (GRPO). Trains immediately when a batch is full. Optional OPD for teacher distillation. | |
madmax |
✅ | Skills + RL + smart scheduler. RL weight updates only run during sleep/idle/meeting windows. |
Serving, reward modeling, and training are fully decoupled. The agent continues responding while scoring and optimization run in parallel.
pip install -e . # skills_only mode (lightweight)
pip install -e ".[rl]" # + RL training support (torch, transformers, tinker)
pip install -e ".[evolve]" # + skill evolution via OpenAI-compatible LLM
pip install -e ".[scheduler]" # + Google Calendar integration for scheduler
pip install -e ".[rl,evolve,scheduler]" # recommended for full RL + scheduler setupIf you want to run rl.backend=mint, install the MinT compatibility package separately in the same environment, for example mindlab-toolkit. MetaClaw keeps that dependency out of the default package so RL users can choose Tinker or MinT explicitly.
metaclaw setupThe interactive wizard will ask you to choose your LLM provider (Kimi, Qwen, MiniMax, or custom), enter your API key, and optionally enable RL training.
MetaClaw's RL path can switch explicitly between tinker and mint. auto is the recommended default and will still infer MinT from Mint-like credentials or base URLs when the MinT package is installed.
Tinker:
metaclaw config rl.backend tinker
metaclaw config rl.api_key sk-...
metaclaw config rl.model moonshotai/Kimi-K2.5MinT:
metaclaw config rl.backend mint
metaclaw config rl.api_key sk-mint-...
metaclaw config rl.base_url https://mint.macaron.xin/
metaclaw config rl.model Qwen/Qwen3-4B-Instruct-2507Legacy aliases rl.tinker_api_key and rl.tinker_base_url are still accepted for backward compatibility.
metaclaw startThat's it. MetaClaw starts the proxy, automatically configures OpenClaw to use it, and restarts the gateway. Open OpenClaw and start chatting — skills are injected at every turn, and the session is automatically summarized into new skills when you're done.
Configuration lives in ~/.metaclaw/config.yaml, created by metaclaw setup.
CLI commands:
metaclaw setup # Interactive first-time configuration wizard
metaclaw start # Start MetaClaw (default: madmax mode)
metaclaw start --mode rl # Force RL mode (no scheduler) for this session
metaclaw start --mode skills_only # Force skills-only mode for this session
metaclaw stop # Stop a running MetaClaw instance
metaclaw status # Check proxy health, running mode, and scheduler state
metaclaw config show # View current configuration
metaclaw config KEY VALUE # Set a config value
Full config reference (click to expand)
mode: madmax # "madmax" | "rl" | "skills_only"
llm:
provider: kimi # kimi | qwen | openai | minimax | custom
model_id: moonshotai/Kimi-K2.5
api_base: https://api.moonshot.cn/v1
api_key: sk-...
proxy:
port: 30000
api_key: "" # optional bearer token for the local MetaClaw proxy
skills:
enabled: true
dir: ~/.metaclaw/skills # your skill library
retrieval_mode: template # template | embedding
top_k: 6
task_specific_top_k: 10 # cap task-specific skills (default 10)
auto_evolve: true # auto-summarize skills after each session
rl:
enabled: false # set to true to enable RL training
backend: auto # "auto" | "tinker" | "mint"
model: moonshotai/Kimi-K2.5
api_key: ""
base_url: "" # optional backend endpoint, e.g. https://mint.macaron.xin/ for MinT
tinker_api_key: "" # legacy alias for api_key
tinker_base_url: "" # legacy alias for base_url
prm_url: https://api.openai.com/v1
prm_model: gpt-5.2
prm_api_key: ""
lora_rank: 32
batch_size: 4
resume_from_ckpt: "" # optional checkpoint path to resume training
evolver_api_base: "" # leave empty to reuse llm.api_base
evolver_api_key: ""
evolver_model: gpt-5.2
opd:
enabled: false # set to true to enable OPD (teacher distillation)
teacher_url: "" # teacher model base URL (OpenAI-compatible /v1/completions)
teacher_model: "" # teacher model name (e.g., Qwen/Qwen3-32B)
teacher_api_key: "" # teacher model API key
kl_penalty_coef: 1.0 # KL penalty coefficient for OPD
max_context_tokens: 20000 # prompt token cap before truncation
scheduler: # v0.3: meta-learning scheduler (auto-enabled in madmax mode)
enabled: false # madmax mode enables this automatically; set manually for rl mode
sleep_start: "23:00"
sleep_end: "07:00"
idle_threshold_minutes: 30
min_window_minutes: 15
calendar:
enabled: false
credentials_path: ""
token_path: ""metaclaw start --mode skills_only
The lightest mode. No GPU, no RL backend needed. MetaClaw places your LLM behind a proxy that injects relevant skills at every turn, then auto-summarizes new skills after each conversation.
Skills are short Markdown instructions stored in ~/.metaclaw/skills/ as individual SKILL.md files. The library grows automatically with your usage.
To pre-load the built-in skill bank (40+ skills across coding, security, agentic tasks, etc.):
cp -r memory_data/skills/* ~/.metaclaw/skills/metaclaw start --mode rl
Everything in Skills Mode, plus continuous RL fine-tuning from live conversations. Each conversation turn is tokenized and submitted as a training sample. A judge LLM (PRM) scores responses asynchronously, and a Tinker-compatible backend (Tinker cloud or MinT) runs LoRA fine-tuning with hot-swapped weights.
Tinker:
metaclaw config rl.backend tinker
metaclaw config rl.api_key sk-...
metaclaw config rl.model moonshotai/Kimi-K2.5
metaclaw config rl.prm_url https://api.openai.com/v1
metaclaw config rl.prm_api_key sk-...
metaclaw start --mode rlMinT:
metaclaw config rl.backend mint
metaclaw config rl.api_key sk-mint-...
metaclaw config rl.base_url https://mint.macaron.xin/
metaclaw config rl.model Qwen/Qwen3-4B-Instruct-2507
metaclaw config rl.prm_url https://api.openai.com/v1
metaclaw config rl.prm_api_key sk-...
metaclaw start --mode rlA dedicated evolver LLM also extracts new skills from failed episodes, feeding them back into the skill library.
Programmatic rollout (no OpenClaw TUI needed): set openclaw_env_data_dir to a directory of JSONL task files:
{"task_id": "task_1", "instruction": "Register the webhook at https://example.com/hook"}OPD is an optional add-on for RL Mode. It distills a larger teacher model into the student on-policy: the student generates responses as usual, and the teacher provides per-token log-probabilities on those same responses. A KL penalty steers the student toward the teacher's distribution.
metaclaw config opd.enabled true
metaclaw config opd.teacher_url http://localhost:8082/v1
metaclaw config opd.teacher_model Qwen/Qwen3-32B
metaclaw config opd.kl_penalty_coef 1.0The teacher must be served behind an OpenAI-compatible /v1/completions endpoint (e.g., vLLM, SGLang). OPD can be combined with PRM scoring, both run asynchronously. See examples/run_conversation_opd.py and scripts/run_openclaw_tinker_opd.sh.
metaclaw start
Everything in RL Mode, plus a meta-learning scheduler that defers weight updates to user-inactive windows so the agent is never interrupted during active use. This is the default mode.
The RL weight hot-swap step pauses the agent for several minutes. Instead of training immediately when a batch is full (like RL Mode does), MadMax waits for an appropriate window.
Three conditions trigger an update window (any one is sufficient):
- Sleep hours: configurable start/end time (e.g., 23:00 to 07:00)
- Keyboard inactivity: triggers after N minutes of idle time
- Google Calendar events: detects meetings so updates can run while you're away
metaclaw config scheduler.sleep_start "23:00"
metaclaw config scheduler.sleep_end "07:00"
metaclaw config scheduler.idle_threshold_minutes 30
# Optional: Google Calendar integration
pip install -e ".[scheduler]"
metaclaw config scheduler.calendar.enabled true
metaclaw config scheduler.calendar.credentials_path ~/.metaclaw/client_secrets.jsonIf the user returns mid-update, the partial batch is saved and resumed at the next window.
Each ConversationSample is tagged with a skill_generation version. When skill evolution bumps the generation, the RL buffer is flushed so only post-evolution samples are used for gradient updates (MAML support/query set separation).
@misc{xia2026metaclaw,
author = {Xia, Peng and Chen, Jianwen and Yang, Xinyu and Tu, Haoqin and Han, Siwei and Qiu, Shi and Zheng, Zeyu and Xie, Cihang and Yao, Huaxiu},
title = {MetaClaw: Just Talk --- An Agent That Meta-Learns and Evolves in the Wild},
year = {2026},
organization = {GitHub},
url = {https://github.com/aiming-lab/MetaClaw},
}MetaClaw builds on top of the following open-source projects:
- OpenClaw – the core agent framework.
- SkillRL – our skill-augmented RL framework.
- Tinker – used for online RL training.
- MinT – alternative backend for online RL training.
- OpenClaw-RL – inspiration for our RL design.
- awesome-openclaw-skills – provides the foundation for our skill bank.
This project is licensed under the MIT License.