evoclaw 0.2.0

Self-evolving AI agents via LoRA — Just talk to your agent, it learns.

🦎 EvoClaw

Just talk to your agent — it learns and EVOLVES.

EvoClaw turns live conversations into continuous training data — automatically.
Works with any OpenAI-compatible API. Uses free Groq for PRM scoring. Trains with Tinker cloud LoRA.

---

🔥 What is EvoClaw?

EvoClaw wraps your existing AI agent behind an OpenAI-compatible proxy. Every conversation is:

1. Scored by a PRM (Process Reward Model) via Groq
2. Skills extracted from high-quality responses and stored
3. Skills injected into future prompts (immediate improvement, no retraining needed)
4. Failed turns trigger automatic skill evolution via LLM
5. All turns feed Tinker LoRA training (GRPO or OPD)

After every batch_size samples, updated weights are saved to Tinker — no service interruption.

---

🚀 Quick Start

pip install evoclaw

evoclaw init   # enter your Groq + Tinker API keys
evoclaw start  # proxy starts on localhost:8080

Then point your existing OpenAI client at EvoClaw:

from openai import OpenAI

client = OpenAI(
    base_url="http://localhost:8080/v1",
    api_key="any-string",  # Not checked by proxy
)

# Just use it normally — EvoClaw learns in the background
response = client.chat.completions.create(
    model="llama-3.3-70b-versatile",
    messages=[{"role": "user", "content": "Explain impermanent loss"}]
)

That's it. Start chatting. EvoClaw learns automatically.

---

🤖 Key Features

Skill Injection

At every turn, the most relevant learned skills are injected into the system prompt.
Immediate behavior improvement — no waiting for retraining.

Skill Evolution

When the agent fails (low PRM score), EvoClaw uses an LLM to generate a new skill
that would have prevented the failure. Over time, the skill bank grows smarter.

Tinker LoRA Training

All conversations feed into online LoRA training via Tinker.
No GPU required. Updated weights are hot-swapped with no downtime.

Two Learning Modes

• GRPO: Reinforcement learning from implicit conversation rewards
• OPD: On-policy distillation from high-quality responses

Works with Any Provider

Unlike MetaClaw (OpenClaw + Kimi-2.5 only), EvoClaw works with:

• Groq (free, recommended)
• OpenAI
• Anthropic
• Any OpenAI-compatible endpoint

---

⚙️ Configuration

All settings in EvoClawConfig:

Field	Default	Description
model_name	Qwen/Qwen3-4B	Tinker base model
lora_rank	32	LoRA rank
batch_size	32	Samples before train step
loss_fn	importance_sampling	grpo / opd / cross_entropy
use_prm	True	PRM scoring
prm_threshold	0.65	Min score to learn from
use_skills	True	Skill injection
enable_skill_evolution	True	Auto-generate skills from failures
proxy_port	8080	Proxy listen port

---

💪 Skill Packs

Pre-built skills for common domains:

config = EvoClawConfig(
    skill_packs=["general", "coding", "crypto", "defi", "security", "agentic"]
)

---

🔄 Training Loop Example

python examples/run_conversation_rl.py           # GRPO mode
python examples/run_conversation_rl.py --mode opd  # OPD mode
python examples/run_conversation_rl.py --no-train  # Skill injection only

Train from your own conversation file:

evoclaw train --file conversations.jsonl
# Format: {"user": "...", "assistant": "..."}

---

📊 Monitor Progress

evoclaw status        # Skills + trainer status
evoclaw skills        # List all learned skills  
evoclaw skills --category crypto  # Filter by category

---

🏗️ Architecture

User/Agent
    │
    ▼
┌─────────────────────────────────┐
│  EvoClaw Proxy (localhost:8080) │
│  - Inject skills into prompt    │
│  - Forward to upstream API      │
│  - Score response async (Groq)  │
│  - Evolve skills on failure     │
│  - Feed samples to Tinker       │
└─────────────────────────────────┘
    │              │
    ▼              ▼
Groq API     Tinker LoRA
(responses)  (training)

---

📄 License

MIT

Acknowledgements

Built on top of MetaClaw, Tinker, and Groq.