Cocoon Skill for OpenClaw

An OpenClaw skill that gives your AI agent access to Cocoon — confidential AI inference running in trusted execution environments on the TON blockchain.

What is Cocoon?

COCOON (Confidential Compute Open Network) is a decentralized AI inference platform where:

• GPU owners earn TON by serving models inside Intel TDX enclaves
• Developers get low-cost, secure, and verifiable AI compute
• Users get AI with full privacy — requests and responses stay encrypted end-to-end

Models run inside hardware-attested trusted execution environments. Nobody — not the GPU operator, not the proxy, not anyone in between — can see your prompts or completions.

What This Skill Does

This skill wraps the Cocoon client's OpenAI-compatible API into simple commands your OpenClaw agent can use:

• Health checks — Verify the Cocoon client is running
• Model discovery — List available models on the network
• Chat completions — Send messages, get responses
• Streaming — Real-time token-by-token output
• Text completions — Raw prompt completions
• Stats — Monitor usage and performance

Why Use This?

Without This Skill	With This Skill
Hand-craft curl to localhost:10000/v1/chat/completions	cocoon.sh chat "Hello"
Manually discover model names	Auto-detected from /v1/models
No streaming without careful flag management	cocoon.sh stream "Write a poem"
Parse raw JSON responses yourself	Structured output for agent consumption
Forget Cocoon-specific params	--max-coefficient, --timeout, --debug built in

Architecture

┌──────────────┐     ┌──────────────────┐     ┌───────────┐     ┌──────────────────────┐
│  OpenClaw    │────▶│  Cocoon Client   │────▶│   Proxy   │────▶│  Worker (Intel TDX)  │
│  Agent       │     │  localhost:10000  │     │  RA-TLS   │     │  vLLM / SGLang + GPU │
└──────────────┘     └──────────────────┘     └───────────┘     └──────────────────────┘
     asks                 forwards              routes             runs model in TEE
     "chat ..."           via OpenAI API        to workers         returns completion

All traffic between Proxy and Worker is encrypted via RA-TLS (Remote Attestation TLS). The Worker's TEE image hash is verified on-chain before any request is routed to it.

Quick Start

# 1. Install the skill
cd ~/oni/aton/logan/workspace/skills
git clone https://github.com/AlphaTONCapital/cocoon-claw-skill.git cocoon

# 2. Ensure a Cocoon client is running (default: localhost:10000)
#    Or set a custom endpoint:
export COCOON_ENDPOINT="http://your-cocoon-client:10000"

# 3. Check health
./cocoon/scripts/cocoon.sh health

# 4. Send your first request
./cocoon/scripts/cocoon.sh chat "What is confidential computing?"

Usage

For OpenClaw Agents

Once installed, your agent picks up the skill automatically:

You: "Ask Cocoon what models are available"
Agent: [Uses cocoon.sh models to list]

You: "Use Cocoon to summarize this document"
Agent: [Uses cocoon.sh chat with the document content]

You: "Stream a response from Cocoon about TON"
Agent: [Uses cocoon.sh stream for real-time output]

Command Line Interface

# Health & status
./scripts/cocoon.sh health                    # Check if client is running
./scripts/cocoon.sh models                    # List available models
./scripts/cocoon.sh stats                     # Get JSON usage stats

# Inference
./scripts/cocoon.sh chat "message"            # Chat completion
./scripts/cocoon.sh stream "message"          # Streaming chat completion
./scripts/cocoon.sh complete "prompt"         # Text completion

# With options
./scripts/cocoon.sh chat "Hello" --model Qwen/Qwen3-8B --max-tokens 200 --temperature 0.5
./scripts/cocoon.sh stream "Write a haiku" --timeout 30 --debug

Examples

# Quick chat
./scripts/cocoon.sh chat "Explain zero-knowledge proofs in one sentence"

# Stream a longer response
./scripts/cocoon.sh stream "Write a short story about a lobster on the blockchain" \
  --max-tokens 1000 --temperature 0.9

# Text completion with a specific model
./scripts/cocoon.sh complete "The future of decentralized AI is" \
  --model Qwen/Qwen3-8B --max-tokens 256

Features

• Minimal dependencies — Requires only curl. Uses jq or python3 for JSON escaping when available, falls back to sed
• Auto model detection — Discovers available models automatically
• Streaming support — Token-by-token output via SSE
• Cocoon-specific params — --max-coefficient, --timeout, --debug
• Configurable endpoint — COCOON_ENDPOINT env var
• OpenAI-compatible — Standard /v1/chat/completions API under the hood

Repository Structure

cocoon-claw-skill/
├── README.md              # This file
├── LICENSE                # MIT License
├── SKILL.md               # OpenClaw skill definition
├── scripts/
│   ├── cocoon.sh          # CLI tool for all Cocoon operations
│   └── test.sh            # Test suite
└── references/
    └── api.md             # Complete Cocoon API reference

How It Works

1. OpenClaw loads SKILL.md when you mention Cocoon or confidential inference
2. Skill provides context — API endpoints, usage patterns, available commands
3. Agent calls scripts/cocoon.sh with the appropriate command
4. Script sends HTTP requests to the local Cocoon client (localhost:10000)
5. Cocoon client routes through Proxy to a TEE-protected Worker, returns the result

Configuration

Variable	Default	Description
COCOON_ENDPOINT	http://127.0.0.1:10000	Cocoon client HTTP endpoint

All configuration is via environment variables. No credentials are stored — the Cocoon client handles authentication and payment channels with the TON blockchain.

Security

• Trusted Execution — All inference runs inside Intel TDX hardware enclaves
• RA-TLS — Proxy-to-Worker connections are encrypted and attestation-verified
• On-chain verification — Worker image hashes are registered in TON smart contracts
• Local only — This skill talks to your local Cocoon client; nothing leaves your machine until the client encrypts it
• No credentials stored — No API keys, tokens, or secrets in this repo

Troubleshooting

"Cocoon client not responding"

# Check if the client process is running
curl -s http://127.0.0.1:10000/stats

# Verify the endpoint
echo $COCOON_ENDPOINT

"No model available"

• Ensure the Cocoon client is connected to at least one proxy with active workers
• Check ./scripts/cocoon.sh models — if empty, no workers are serving models yet
• Try specifying a model explicitly: --model Qwen/Qwen3-8B

"504 Gateway Timeout"

• No workers available on the network, or request timed out
• Try increasing timeout: --timeout 60
• Check network stats: ./scripts/cocoon.sh stats

Contributing

Contributions welcome. Open an issue or submit a PR.

1. Fork the repository
2. Create a feature branch
3. Make your changes
4. Submit a pull request

License

MIT — See LICENSE file.

Links

• Cocoon: https://github.com/TelegramMessenger/cocoon
• OpenClaw: https://openclaw.ai
• TON: https://ton.org
• This Repo: https://github.com/AlphaTONCapital/cocoon-claw-skill

---

Built by AlphaTON Capital — bridging AI agents and confidential compute.